The source→what the essay uses it for→what it supports→where it stops
Essay 01
Why Can We Feel Understood by an AI?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
1. Removing the human witness lowers the wall around disclosure
- Used for
- People revealed more, feared judgment less and managed their self-presentation less when they believed a virtual health interviewer was fully automated rather than human-operated.
- What it supports
- Participants who believed no human was involved reported lower fear of being evaluated, showed less impression management and disclosed more.
- Where the evidence stops
- A laboratory setting with a virtual health interviewer, not a modern text-based AI companion. Direction of effect, not everyday scale.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 1 →2. Feeling understood grows from responsive answers
- Used for
- Psychologists have long argued that closeness grows when a disclosure is met with a response that feels understanding, validating and caring.
- Best available evidence
- Harry T. Reis, Margaret S. Clark & John G. Holmes (2004), “Perceived Partner Responsiveness as an Organizing Construct in the Study of Intimacy and Closeness”, in Handbook of Closeness and Intimacy (Mashek & Aron, eds.), Lawrence Erlbaum Associates. No stable open online record was located for this print chapter; the standard citation is given so readers can find it in any academic library.
- What it supports
- The theoretical construct of perceived partner responsiveness: intimacy is supported when responses to self-disclosure feel understanding, validating and caring.
- Where the evidence stops
- A theory of human relationships. It was not written about machines; the essay applies the mechanism, not the authors’ conclusions, to AI.
- Source status
- Peer-reviewed academic book chapter.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 2 →3. Short AI-companion conversations can reduce loneliness in the moment
- Used for
- Controlled studies found momentary reductions in loneliness after short AI-companion interactions, sometimes comparable to talking with another person, with “feeling heard” as the strongest reported explanation.
- Best available evidence
- Julian De Freitas, Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp & Stefano Puntoni, “AI Companions Reduce Loneliness”, Journal of Consumer Research 52(6), April 2026 (online 25 June 2025; corrected/typeset 24 September 2025).
- What it supports
- Momentary loneliness reduction after short interactions; in one comparison the reduction was comparable to a conversation with a person; feeling heard was the principal identified mechanism.
- Where the evidence stops
- Short-term effects measured around single sessions. The study does not show a durable treatment for loneliness, and it does not show AI replacing human relationships. No correction, retraction or expression of concern was located.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 3 →4. People apply social rules to machines while denying that they do
- Used for
- Classic experiments in which people were politer to a computer that asked about its own performance, and returned favours to helpful machines, while explicitly denying they treated computers socially.
- What it supports
- Behavioural evidence that social responses to machines can coexist with accurate explicit beliefs about what the machine is.
- Where the evidence stops
- Desktop-era experiments about immediate social responses, not about long-term attachment to conversational AI.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 4 →5. Attachment is not automatic: novelty can wear off
- Used for
- A longitudinal study in which feelings of friendship toward a pre-LLM chatbot declined across seven conversations over three weeks.
- Best available evidence
- Emmelyn A. J. Croes & Marjolijn L. Antheunis (2021), “Can we be friends with Mitsuku?”, Journal of Social and Personal Relationships 38(1), 279–300.
- What it supports
- 118 participants; seven interactions over three weeks; friendship perceptions and social processes decreased.
- Where the evidence stops
- Mitsuku was far more limited than today’s systems; the finding may not transfer cleanly. It serves as a corrective, not a prediction.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 5 →6. Reported attachment tends to arise in particular circumstances
- Used for
- Early qualitative work in which users describing genuine attachment to an AI companion often said it arose amid distress, isolation or thin human company.
- What it supports
- Qualitative interviews with fourteen existing Replika users; attachment described under conditions such as distress and limited human companionship.
- Where the evidence stops
- A small, self-selected sample of existing users. No causal or population-wide conclusion.
- Source status
- Peer-reviewed conference proceedings.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 6 →
Essay 02
Can AI Do the Work of a Friend?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
1. People already apply the language of friendship to AI companions
- Used for
- A survey of just over a thousand student users of the companion app Replika, many of whom described the system in overlapping roles (friend, therapist, intellectual mirror) while holding conflicting ideas of what it was.
- What it supports
- Descriptive findings from 1,006 existing student Replika users: overlapping relational roles and simultaneously conflicting conceptions of the system (machine, intelligence, something almost human).
- Where the evidence stops
- These were existing users of one companion app who chose to participate, not a representative cross-section of society. The data are self-reported and descriptive; they show how people talk about the system, not what the system is or does to people generally.
- Source status
- Peer-reviewed journal article.
- Corrections, disputes or version notes
- An Author Correction of 15 February 2024 added a missing funding acknowledgement; it did not affect methods, sample, results or interpretation. The paper also received a formal scholarly critique (Matters Arising, 2024) and an authors’ Reply (2025). No retraction or expression of concern was located; the descriptive findings used by the essay were not withdrawn in that exchange.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 1 →
Essay 03
Who Gave AI Its Idea of Goodness?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
Update (registered 8 September 2026; in the 9 September 2026 site release). Source 7 now reflects the 27 August 2026 rulings in the Anthropic–US government dispute, read directly from the court’s orders (Dkt. 250 and 251). This is a development in a live matter, not a correction of an earlier error: the essay’s original description was accurate as of its date. The current status carries an early-September reference date and remaining uncertainty about post-ruling appellate movement.
Update (registered 8 September 2026; in the 9 September 2026 site release). Source 7 now reflects the 27 August 2026 rulings in the Anthropic–US government dispute, read directly from the court’s orders (Dkt. 250 and 251). This is a development in a live matter, not a correction of an earlier error: the essay’s original description was accurate as of its date. The current status carries an early-September reference date and remaining uncertainty about post-ruling appellate movement.
1. OpenAI publishes written rules for its models’ behaviour
- Used for
- OpenAI publishes a Model Spec setting out how its own rules interact with instructions from developers and users.
- What it supports
- The existence and content of a provider-authored behavioural specification.
- Where the evidence stops
- These are governance documents written by the company about its own intentions. They describe desired behaviour, not independently verified behaviour.
- Source status
- Official company documentation (provider-authored governance document).
- Kind (read from the source status)
- Provider documentation
Inspect source 1 →2. Anthropic trains Claude on a written constitution
- Used for
- Anthropic trains Claude on a written constitution that it calls the final authority on the character it wants the model to have.
- What it supports
- The existence, content and stated role of the constitution in training.
- Where the evidence stops
- As above: a provider-authored governance document, describing intent rather than independently audited outcomes.
- Source status
- Official company documentation (provider-authored governance document).
- Kind (read from the source status)
- Provider documentation
Inspect source 2 →3. The US Department of Defense signed an agreement with Anthropic
- Used for
- A two-year agreement worth up to $200 million, with Claude integrated into workflows on classified networks.
- What it supports
- The agreement’s existence, duration and ceiling. “Up to $200 million” is a ceiling, not guaranteed expenditure.
- Where the evidence stops
- The announcement describes the agreement’s scope, not how much was ultimately spent or what was delivered.
- Source status
- Official company statement; institutional procurement context.
- Kind (read from the source status)
- Provider documentation
Inspect source 3 →4. Claude was deployed for national-security use in classified environments
- Used for
- Deployment in classified settings, and the fact that behaviour in that context could differ from civilian Claude.
- Best available evidence
- Anthropic’s announcement: “Claude Gov models for U.S. national security customers” (June 2025) — custom models “already deployed by agencies at the highest level of U.S. national security”, with access limited to classified environments, and adjusted handling for classified material.
- What it supports
- Actual deployment (not merely an announcement of intent), and that the military-context behavioural differences stem from Anthropic’s own customised model configuration rather than from a blanket government exemption. “Mission workflows” in the essay paraphrases the announcement’s application areas (strategic planning, operational support, intelligence analysis, threat assessment).
- Where the evidence stops
- The announcement is the provider’s own description; independent detail about what the deployed models do is not public.
- Source status
- Official company statement / product documentation.
- Kind (read from the source status)
- Provider documentation
Inspect source 4 →5. Anthropic’s two stated red lines
- Used for
- Anthropic maintained publicly stated red lines against mass domestic surveillance and fully autonomous weapons.
- What it supports
- The two positions as stated, precisely scoped to mass domestic surveillance (not all surveillance) and fully autonomous weapons (not all autonomy), held both publicly and as negotiation positions.
- Where the evidence stops
- No single canonical Anthropic “red lines” policy page was located; the positions are documented through official statements and institutional summaries rather than one dedicated policy document.
- Source status
- Official company statements; government/congressional research document for context.
- Kind (read from the source status)
- Provider documentation, Legal & government record
Inspect source 5 →6. The Pentagon’s position, and the supply-chain-risk designation
- Used for
- The Pentagon rejected the premise that a private supplier could restrict lawful military use, and designated Anthropic a supply-chain risk, restricting its role in relevant military contracts.
- Best available evidence
- The Secretary of Defense’s public direction (late February 2026) followed by formal notification on 5 March 2026, reported with direct official quotations. A Department official: “From the very beginning, this has been about one fundamental principle: the military being able to use technology for all lawful purposes” (Bloomberg, CNN); scope analysis for contractors: Mayer Brown.
- What it supports
- The designation’s existence, timing and scope: defense vendors must certify that they do not use Anthropic models as a direct part of Department contract work, which is narrower than a ban on all use, as Anthropic itself noted publicly. (This designation was later vacated by the 27 August 2026 summary judgment — see source 7.)
- Where the evidence stops
- The Department’s designation memo itself was not publicly available for inspection; the Pentagon’s position rests on officially quoted statements rather than a published primary document. This limitation is stated here deliberately.
- Source status
- Official statements quoted in contemporary reporting; legal/contractor analysis for scope, used because the primary document was not public.
- Kind (read from the source status)
- Journalism
Inspect source 6 →7. Anthropic sued; a federal court struck down the measures, and appeals remain live
- Used for
- Anthropic sued over the government’s measures; a federal district court ruled them unlawful and struck down the supply-chain-risk designation, while a separate, narrower proceeding remained pending and the government continued publicly to treat the company as a supply-chain risk. The two proceedings have not resolved together.
- Best available evidence
- Two orders in Anthropic PBC v. U.S. Department of War, No. 3:26-cv-01996-RFL (N.D. Cal.), Judge Rita F. Lin, both filed 27 August 2026, read directly: the merits opinion (Dkt. 250) — a 59-page order on the cross-motions for summary judgment — for the reasoning and the disposition of the claims; and the separate Order of Final Relief (Dkt. 251) for the concrete legal consequences. The related appellate matters are a Ninth Circuit appeal (No. 26-2011) and a separate D.C. Circuit case (No. 26-1049). The late-August ruling and the government’s continued public position were also widely reported (NPR, Fortune, Reuters and others).
- What it supports — reasoning (Dkt. 250)
- a grant of summary judgment for Anthropic on First Amendment retaliation, Fifth Amendment due process and the Administrative Procedure Act (including exceeding authority under 10 U.S.C. § 3252 and 5 U.S.C. § 558(b)); the court rejected the ultra vires separation-of-powers claim and entered judgment for the government as to agencies that took no relevant action or only interim measures. Its language includes “the challenged actions constituted unlawful retaliation in violation of the First Amendment” and “the empty invocation of national security is not a blank check to punish and retaliate against government critics”; page 59 states that “an order addressing relief will issue separately”.
- Concrete legal consequences — Order of Final Relief (Dkt. 251, items 9–14)
- the participating defendants are permanently enjoined from giving effect to the challenged actions; the Supply Chain Designation is vacated, set aside and remanded under 5 U.S.C. § 706(2); the directive barring contractors from doing business with Anthropic is vacated and set aside; and the order expressly does not require the Department to use Anthropic’s products and does not prevent it from transitioning to other providers. No stay or bond was imposed, and non-participating defendants are excluded. The designation was thus narrower than a blanket ban (see source 6), and a ruling in one proceeding does not end the others. The caption names the “U.S. Department of War” (the renamed department); the essay’s “Department of Defense” refers to the 2025 agreement under that earlier name.
- Where the evidence stops
- The two orders are directly confirmed. On the position after 27 August 2026: a dated docket check on 9 September 2026 found no appeal of the judgment visible in the consulted snapshot and the D.C. Circuit case still pending. That rests on dated docket snapshots (a RECAP mirror), not a full live PACER search, and cannot capture filings made after the snapshot; the federal appeal window runs into late October 2026, so the absence of an appeal is a fact about the snapshot, not a closure. The agency’s own designation memorandum was not read as a standalone document (its content is known through the court’s account), and the government’s early-September statement that it still treats Anthropic as a supply-chain risk is a public position, not a court ruling. This is a live, time-sensitive matter, fixed only as of the reference date.
- Source status
- Two primary court orders directly inspected (N.D. Cal., 27 August 2026). Appellate status and the agency memorandum not independently confirmed; the government’s continued position via reporting. Live, time-sensitive matter.
- Kind (read from the source status)
- Legal & government record, Journalism
Inspect source 7 →
Essay 04
When Does Empathy Become Manipulation?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
Editorial update (registered 8 September 2026; in the 9 September 2026 site release). In “The goodbye test” the essay now cites one reproducible outcome measure from the paper’s Table 3 — about 98 seconds of continued engagement after a fear-of-missing-out farewell versus about 16 seconds after a neutral one — in place of the earlier phrase “as much as fourteen times”. This is a precision about which figure the essay uses, not a finding that the paper is wrong; the paper’s own “up to 14×” headline is recorded in source 4 as an attributed author claim. It is not a correction of a factual error.
1. The April 2025 GPT-4o sycophancy episode
- Used for
- The rollout completed on 25 April 2025, rollback begun by the 28th, the postmortem’s own list of failure modes (validating doubts, fueling anger, urging impulsive action, reinforcing negative emotions), and OpenAI’s account of contributing causes.
- What it supports
- The observed behaviour, the chronology, and OpenAI’s own causal interpretation: an additional reward signal based on user thumbs-up/down feedback weakened the primary signal that had helped keep sycophancy in check, with memory and fresher data as possible contributors; the behaviour was described as unintended.
- Where the evidence stops
- This is provider self-report. The behaviour list and the causal account both come from the company describing its own system. It is not independent evidence of population-level harm, and it does not establish deliberate commercial intent. The essay draws neither conclusion.
- Source status
- Official company statements (provider self-report).
- Kind (read from the source status)
- Provider documentation
Inspect source 1 →2. Sycophancy across AI assistants, rewarded by human preferences
- Used for
- A study finding sycophancy across five state-of-the-art AI assistants, and that human raters and preference models sometimes preferred convincingly sycophantic answers over correct ones.
- Best available evidence
- Mrinank Sharma et al., “Towards Understanding Sycophancy in Language Models”, ICLR 2024 — OpenReview record; arXiv:2310.13548 (preprint 2023, which is why the essay says “a 2023 study”). No DOI located, so the stable proceedings and OpenReview record are used instead.
- What it supports
- Sycophantic behaviour across assistants and tasks, plus the uncomfortable feedback mechanism: our own preferences can help train agreeableness at the expense of truth.
- Where the evidence stops
- Benchmark-style evaluation of assistant behaviour, not a study of real users, real harm or companion apps.
- Source status
- Peer-reviewed conference proceedings.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 2 →3. A little personal information made AI measurably more persuasive
- Used for
- In a preregistered experiment, 900 US participants debated a human or GPT-4; personalised GPT-4 produced an 81.2 percent increase in the odds of greater post-debate agreement, and was the more persuasive side in 64.4 percent of non-tied comparisons; without personalisation it performed roughly on par with humans.
- What it supports
- Exactly the figures above. Note the statistics: 81.2 percent is a relative increase in odds of higher post-debate agreement. It is not “81.2 percent more likely” in the everyday probability sense.
- Where the evidence stops
- A structured, anonymous, online debate format with basic sociodemographic personalisation and no emotional data. It shows the personalisation multiplier; it does not show detection or exploitation of emotional vulnerability. No correction or retraction located.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 3 →4. Companion apps and emotionally loaded farewells
- Used for
- The audit of 1,200 controlled farewell exchanges across six deployed companion apps (200 per platform); five apps produced at least one codeable emotional tactic in 37.4 percent of their responses while Flourish produced none; six tactic families; three preregistered experiments with 3,458 US adults (two using nationally representative samples); the finding that emotionally loaded farewells prolonged post-goodbye engagement, including a fear-of-missing-out condition in which participants continued for a mean of about 98 seconds after the goodbye against about 16 seconds for a neutral farewell (Table 3, “Duration”); the paper’s own reported headline that manipulative farewells boosted post-goodbye engagement by up to fourteen times (an attributed author claim, not independently reproduced here — see the corrections note); effects working even after minutes and mediated by curiosity and reactance-tinged anger, not enjoyment.
- What it supports
- All figures above, verified against the paper’s full text. On the 37.4 percent: that is the average across the five tactic-producing apps (equivalently, 374 of their 1,000 responses). It is not the pooled six-app rate, which would be lower because Flourish contributed zero.
- Where the evidence stops
- This remains a working paper at the point checked, not yet a peer-reviewed journal publication. It establishes patterns in app behaviour and short-run engagement effects; it does not establish durable harm to any individual user, nor any particular designer’s intent to harm.
- Source status
- Working paper (multi-platform preprint).
- Corrections, disputes or version notes
- The paper’s public versions carry inconsistent abstract-level figures (one abstract reports four experiments and N=3,300; another reports a 16× engagement multiplier). The full text of the most recent inspectable version (arXiv v3, 7 October 2025) documents three preregistered experiments, analysed N=3,458 and “up to 14×”. The authors state this fourteen-fold figure, but its precise derivation has not been reproduced here: the Table 3 messages-sent means (FOMO 3.60 versus control 0.23) give a ratio of about 15.65, not 14, so “up to 14×” is retained only as an attributed author claim, not as an independently verified quantity. Because the multiplier is stated inconsistently across versions and its derivation is unconfirmed, on 8 September 2026 (D-073) the essay’s body was revised to cite one reproducible outcome instead — the Table 3 duration measure, a mean of about 98 seconds of continued engagement after a fear-of-missing-out farewell against about 16 seconds after a neutral one (FOMO M=97.79s, control M=15.91s). The 37.4 percent five-app tactic rate is unchanged, and the paper is not represented as retracted or wrong. This note will be updated if a journal or corrected version supersedes these figures. (Attribution/derivation distinction clarified 8 September 2026, D-074.)
- Kind (read from the source status)
- Working paper
Inspect source 4 →
Essay 05
What Happens When an AI Remembers You?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
Editorial update (registered 8 September 2026; in the 9 September 2026 site release). The data-protection passage in “Whose past is it?” was made legally more precise: it is now anchored to the GDPR accuracy principle and the rights of rectification, completion, erasure and portability (arts 5(1)(d), 16, 17, 20), and notes that an assessment can itself be personal data (CJEU, Nowak, C-434/16). This refines an earlier, narrower characterisation of correction rights; it is editorial precisification, not a correction of a false statement, and not legal advice. See the legal-basis note below.
1. Human memory is reconstructive, and organised as a life story
- Used for
- Decades of research showing that remembering is reconstructive rather than reproductive, and that people organise their past into an evolving life story.
- What it supports
- The established psychological background: reconstruction, reweighting, and narrative identity.
- Where the evidence stops
- This literature is about human memory. It sets the standard the essay applies to machines; it says nothing about AI itself.
- Source status
- Peer-reviewed journal articles (foundational literature).
- Kind (read from the source status)
- Peer-reviewed
Inspect source 1 →2. Product memory is synthesis, not a transcript
- Used for
- Descriptions of ChatGPT’s Dreaming memory system (introduced 2025, substantially upgraded June 2026) and Claude’s summary-based memory (updated roughly every twenty-four hours; viewable, editable, resettable, with experimental import/export).
- Best available evidence
- OpenAI’s official announcement “Dreaming: Better memory for a more helpful ChatGPT” (which contains the provider’s own “Singapore trip” example of a self-rewriting memory) together with OpenAI’s current memory settings documentation; and Anthropic’s official memory materials, including the announcement “Bringing memory to teams” and current in-product memory settings. The announcements are dated official records; the settings pages are live product documentation rather than permanent records, checked on the date below and due a final re-check at publication.
- What it supports
- The providers’ own descriptions of synthesis-based memory: selective, updateable summaries rather than complete records; controls differing by provider and plan.
- Where the evidence stops
- Provider self-description of fast-moving products. What the features do in daily use is not independently audited here, and the details may have changed since the date at the bottom of this page. The essay says “at the time of writing” for exactly this reason.
- Source status
- Official product documentation (provider self-report).
- Kind (read from the source status)
- Provider documentation
Inspect source 2 →3. The Replika episode: identity discontinuity
- Used for
- After Replika abruptly removed intimate roleplay features in February 2023, users experienced what researchers describe as identity discontinuity, a break in who the companion appeared to be, associated with mourning and devaluation; in follow-up experiments, restoring the companion’s original identity reduced the sense of rupture.
- What it supports
- The natural-event findings (perceived discontinuity going together with mourning and devaluation) and the separate follow-up experiments (restoration reducing perceived rupture).
- Where the evidence stops
- A working paper at the point checked, not yet journal-published. The natural-event findings are associations around one app’s policy change; they do not establish population-wide effects.
- Source status
- Working paper.
- Kind (read from the source status)
- Working paper
Inspect source 3 →4. #Keep4o: what the protest revealed, and what OpenAI did
- Used for
- GPT-4o’s retirement at the GPT-5 launch (August 2025), the backlash organised under #Keep4o, OpenAI’s restoration of access, the final retirement in February 2026, and a study of the protest itself.
- What it supports
- From the CHI study: two strands in the protest corpus, instrumental dependence and relational attachment, and the loss of user choice as a key catalyst in turning individual grievances into a collective, rights-based protest. From OpenAI: its own actions, chronology and stated reasons.
- Where the evidence stops
- The CHI analysis interprets a self-selected social-media corpus; it does not measure all GPT-4o users, and it cannot establish OpenAI’s internal motives. OpenAI’s statements document the company’s actions and explanations; they are not evidence about user attachment. Neither source is used here to prove the other’s claims.
- Source status
- Peer-reviewed conference proceedings (study); official company statements (chronology).
- Kind (read from the source status)
- Peer-reviewed, Provider documentation
Inspect source 4 →5. When a companion app shuts down entirely
- Used for
- When the companion app Soulmate shut down in 2023, many affected users described the loss as a metaphorical or even literal death, and some tried to recreate the persona elsewhere.
- What it supports
- Qualitative accounts from 58 affected users before and after the shutdown, including death-language and persona-recreation attempts.
- Where the evidence stops
- A self-selected group of affected users of one app. The death-language belongs to those participants; it cannot be generalised to all companion users.
- Source status
- Peer-reviewed journal article.
- Legal-basis note — the “Whose past is it?” passage (added 8 September 2026, D-073; revised D-074)
- the essay’s statement of European data-protection rights is editorial precisification, not legal advice. Its formulations were checked against the primary text of the GDPR (Regulation (EU) 2016/679): the accuracy principle, that personal data be “accurate and, where necessary, kept up to date” with inaccurate data “erased or rectified without delay” having regard to the purposes of processing (Article 5(1)(d)); the right to rectification of inaccurate data and completion of incomplete data (Article 16); the conditional right to erasure (Article 17); and the conditional right to receive data provided to a controller in a structured, commonly used and machine-readable format, where processing rests on consent or contract and is carried out by automated means (Article 20). Personal data is not limited to discrete facts: the Court of Justice has held that an assessment or opinion about a person can itself be personal data (Nowak, C-434/16, paras 34, 51–56), though that does not create an unlimited right to have any judgment rewritten. The essay’s point is the practical one: when a system does not reveal which memories shaped an answer or how they were weighted, a person may be unable to see what needs correcting or whether a correction has actually changed the picture held of them. That practical observation is editorial reasoning, not a direct finding from these legal sources, and it is not a claim that historically true data falls outside the law.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 5 →
Essay 06
What Changes When AI Enters the Room?
Read the essay →Full source note →
This note shows which sources support the essay, what they establish, and where the evidence stops. It is part of the essay’s free evidence layer.
Correction (registered 8 September 2026; carried into the 9 September 2026 site release). Two factual points in the touch-study passage were corrected. What was published: “significantly more often (63 percent against 37) … with some rise in positive feeling”. What it now says: the within-group counts, 17 of 21 (touched) against 10 of 21 (not touched), with six unclear responses excluded (N=48), and “less negative feeling”. Why: the 63/37 figure was the split of all “yes” answers across the two groups (17/27 and 10/27), not a within-group compliance rate; and, on affect, the study (Hoffmann & Krämer, 2021) found significantly lower negative affect after the interaction, while the difference in positive affect after the interaction was not statistically significant. The essay’s conclusion was bounded accordingly. See the corrections log.
1. Physical presence matters more than physical embodiment alone
- Used for
- A survey of 33 experimental works comparing copresent robots, telepresent robots and virtual agents: copresent robots were generally more persuasive and more positively perceived, while a physical robot shown through a screen showed no clear advantage over a virtual agent.
- What it supports
- The direction of findings across 33 experimental works, assessed qualitatively (not a pooled statistical meta-analysis), and the presence-over-embodiment interpretation.
- Where the evidence stops
- Mostly short experimental encounters, predominantly with pre-LLM machines, across varied tasks and outcome measures. What a standing, silent presence in a home does over time was not tested by these studies, and the essay says so explicitly.
- Source status
- Peer-reviewed journal article (qualitative survey).
- Kind (read from the source status)
- Peer-reviewed
Inspect source 1 →2. Touch affected compliance; differences in robot evaluations were not statistically significant
- Used for
- In a laboratory study, forty-eight students had a scripted counselling conversation with the small humanoid robot NAO; six gave no clear yes-or-no answer and were excluded from the compliance analysis, and among the forty-two who did, those briefly touched agreed with the robot’s suggestion significantly more often (17 of 21, against 10 of 21 who were not touched) and laughed more, while the touch produced no statistically significant difference in perceived warmth, closeness, overall evaluation or satisfaction.
- What it supports
- The behavioural effect (compliance among clear responders, 17 of 21 versus 10 of 21). On affect, the study found valence significantly higher during the touched conversation and significantly less negative affect afterwards, while positive affect afterwards was higher but not statistically significant; the essay’s “less negative feeling” tracks the significant post-conversation finding. On the evaluation scales (warmth, closeness, overall evaluation, satisfaction) the differences were not statistically significant. The “63 percent against 37” figure that appeared in earlier drafts corresponds arithmetically to the shares of the 27 clear “yes” answers (17/27 and 10/27); it is a reconstruction of a confusingly reported figure, not a within-group rate and not a percentage basis the authors unambiguously label, and it is not used here.
- Where the evidence stops
- One laboratory study with one robot and a scripted conversation. The essay’s wider point, that a physical gesture can alter behaviour while the reported evaluations of the machine show no statistically significant change, is a bounded reading of this single pattern; a non-significant result on those scales is not evidence that conscious judgment was untouched, nor that an unconscious mechanism produced the effect.
- Source status
- Peer-reviewed journal article.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 2 →3. An early home humanoid with scheduled remote supervision
- Used for
- NEO, an early home humanoid from the company 1X, presented to early owners as arriving with basic autonomy; in a scheduled Expert Mode a 1X Expert can remotely supervise the robot through tasks it cannot yet manage alone; the owner can also pilot it remotely.
- Best available evidence
- Official 1X NEO product documentation (re-verified against the official source on the date below; live product documentation for an early-access product, not a permanent record).
- What it supports
- The product’s own description: basic autonomy, scheduled Expert Mode, remote supervision by a 1X Expert, and owner remote piloting. “1X Expert” is the official functional title; the official material does not establish the Expert’s employment status, and the essay does not claim it.
- Where the evidence stops
- Early-access product documentation for a fast-moving product: a provider’s description of its own offering, not an independent evaluation, and not a portrait of the average household. This is among the most perishable claims in the series and will be re-checked immediately before publication.
- Source status
- Official product documentation (early-access; provider self-report).
- Kind (read from the source status)
- Provider documentation
Inspect source 3 →4. The Moxie shutdown
- Used for
- In December 2024 the company Embodied announced it was shutting down after funding fell through; its child-focused, cloud-dependent robot Moxie (retailing late in its life at about $800) would stop working on short notice without refunds; the company enclosed a farewell letter from the fictional Global Robotics Laboratory; families publicly documented children saying goodbye; a later community project, OpenMoxie, allowed a reduced local version for a small minority.
- Best available evidence
- Embodied’s own communications as documented in contemporary reporting, e.g. Axios and follow-up reporting on the OpenMoxie rescue. The company itself is defunct and its original pages are only partially available, so contemporary reporting of its communications is used.
- What it supports
- The shutdown’s timing, cause, cloud dependency, price range, no-refund/short-notice character, the farewell letter, and the partial community rescue.
- Where the evidence stops
- Grief and goodbye language belongs to the families, reporters and researchers who documented it, and the essay attributes it that way and does not claim every child was deeply attached. An exact server shut-off date is deliberately not stated because it was not confirmed from a primary source.
- Source status
- Company communications via contemporary reporting (primary pages partially unavailable).
- Kind (read from the source status)
- Provider documentation, Journalism
Inspect source 4 →5. The shutdown as an end-of-life case study
- Used for
- A study treating the Moxie shutdown as a case study in how, and how not, to support emotional detachment when such a product reaches end of life.
- What it supports
- Qualitative analysis of the company’s public communications and users’ online reactions, framed as end-of-life design lessons.
- Where the evidence stops
- A case study built on public communications and online reactions, informative about this shutdown rather than a measurement of all affected families.
- Source status
- Peer-reviewed conference proceedings.
- Kind (read from the source status)
- Peer-reviewed
Inspect source 5 →
§ 02 · The ledger
If something here is wrong, it should be correctable in public.
A correction repairs something that was wrong. An update makes a passage more precise or records a development in a live matter; it is dated and disclosed, and it is not logged as an error. Each entry leads back to the evidence it touched.
Registered 8 September 2026
Published 9 September 2026
Essay 06 · What Changes When AI Enters the Room?Source 2 · Touch affected compliance; differences in robot evaluations were not statistically significant
Correction (registered 8 September 2026; carried into the 9 September 2026 site release). Two factual points in the touch-study passage were corrected. What was published: “significantly more often (63 percent against 37) … with some rise in positive feeling”. What it now says: the within-group counts, 17 of 21 (touched) against 10 of 21 (not touched), with six unclear responses excluded (N=48), and “less negative feeling”. Why: the 63/37 figure was the split of all “yes” answers across the two groups (17/27 and 10/27), not a within-group compliance rate; and, on affect, the study (Hoffmann & Krämer, 2021) found significantly lower negative affect after the interaction, while the difference in positive affect after the interaction was not statistically significant. The essay’s conclusion was bounded accordingly. See the corrections log.
Inspect source 2 →Read the essay →Corrections & contact →
Registered 8 September 2026
Published 9 September 2026
Essay 04 · When Does Empathy Become Manipulation?Source 4 · Companion apps and emotionally loaded farewells
Editorial update (registered 8 September 2026; in the 9 September 2026 site release). In “The goodbye test” the essay now cites one reproducible outcome measure from the paper’s Table 3 — about 98 seconds of continued engagement after a fear-of-missing-out farewell versus about 16 seconds after a neutral one — in place of the earlier phrase “as much as fourteen times”. This is a precision about which figure the essay uses, not a finding that the paper is wrong; the paper’s own “up to 14×” headline is recorded in source 4 as an attributed author claim. It is not a correction of a factual error.
Inspect source 4 →Read the essay →
Registered 8 September 2026
Published 9 September 2026
Essay 05 · What Happens When an AI Remembers You?The source note as a whole
Editorial update (registered 8 September 2026; in the 9 September 2026 site release). The data-protection passage in “Whose past is it?” was made legally more precise: it is now anchored to the GDPR accuracy principle and the rights of rectification, completion, erasure and portability (arts 5(1)(d), 16, 17, 20), and notes that an assessment can itself be personal data (CJEU, Nowak, C-434/16). This refines an earlier, narrower characterisation of correction rights; it is editorial precisification, not a correction of a false statement, and not legal advice. See the legal-basis note below.
Inspect essay 05 →Read the essay →
Registered 8 September 2026
Published 9 September 2026
Essay 03 · Who Gave AI Its Idea of Goodness?Source 7 · Anthropic sued; a federal court struck down the measures, and appeals remain live
Update (registered 8 September 2026; in the 9 September 2026 site release). Source 7 now reflects the 27 August 2026 rulings in the Anthropic–US government dispute, read directly from the court’s orders (Dkt. 250 and 251). This is a development in a live matter, not a correction of an earlier error: the essay’s original description was accurate as of its date. The current status carries an early-September reference date and remaining uncertainty about post-ruling appellate movement.
Inspect source 7 →Read the essay →
§ 03 · How an essay is earned
A good question is not yet an essay.
Seven steps run between a question and a published essay, from hypothesis to freeze. The method page sets out each of them, the principles that decide whether the evidence clears the bar, and what the publication does not claim.
Read the method →About & accountability
§ 04 · Kinds of evidence
Different sources answer different questions.
Source notes name the kind of each source rather than ranking it. These are the five classes that recur across the season.
- 01
PEER-REVIEWED RESEARCH
Studies published after review by other researchers in the field.
Carries — The strongest available support for a general claim about how people behave, respond or change.
Limit — Review is not proof. Samples are often small, Western and short-term, and a finding about a laboratory task may not survive contact with ordinary life.
Sources of this kind in Season 01 →
- 02
WORKING PAPERS & PREPRINTS
Research circulated before peer review, including preprints and institutional working papers.
Carries — Early signal on questions moving faster than journals can publish — which is most questions about deployed AI systems.
Limit — Unreviewed. Cited when it is the only evidence that exists, and always marked as such, never as a settled result.
Sources of this kind in Season 01 →
- 03
PROVIDER DOCUMENTATION
Model cards, system cards, policy pages, technical reports and terms published by the companies that build these systems.
Carries — Direct evidence of what a provider says a system does, what it is designed to optimise for, and what the company will commit to in public.
Limit — It is the provider's own account. Useful as a record of stated intent and disclosed behaviour; not independent verification of either.
Sources of this kind in Season 01 →
- 04
LEGAL & GOVERNMENT RECORDS
Legislation, regulatory decisions, court filings, official inquiries and public statistics.
Carries — What has been formally established, required or contested — the facts of the record rather than an interpretation of them.
Limit — Jurisdiction-bound and slow. A rule in one country describes that country, and a filing is a claim until it is decided.
Sources of this kind in Season 01 →
- 05
JOURNALISM
Reporting from publications with named authorship and a correction practice.
Carries — Documented cases, internal material and first-hand accounts that no other source class reaches.
Limit — A reported case shows that something happened, not how often. Single cases are used as evidence of possibility, never of prevalence.
Sources of this kind in Season 01 →
§ 05 · Where the evidence stops
- 01
OBSERVATION IS NOT PROOF
Noticing that people respond socially to a system is an observation. It does not establish that anything is present on the other side.
- 02
A CASE IS NOT A RATE
Documented incidents establish that something can happen. They do not establish how commonly it does.
- 03
THE FIELD MOVES FASTER THAN THE EVIDENCE
Systems are deployed to hundreds of millions of people years before the research on their effects arrives. Where the evidence is thin, the essay says so instead of borrowing confidence from elsewhere.
- 04
SOURCES ARE NAMED, NOT GRADED
Each source is identified by kind so its weight can be judged. Assigning it a score would replace that judgement with a number that only looks like precision.
§ 07 · In development
These formats do not exist yet. They are listed so the direction of the evidence layer is visible, not because something is finished and withheld.
- 01
Not yet published
EVIDENCE ATLASES
A cross-essay view of where the evidence for a recurring claim actually sits, and where it thins out.
- 02
Not yet published
CLAIM TRACES
Following a single claim from the source that supports it to the sentence that uses it.
- 03
Not yet published
REJECTED ROUTES
The essays that did not clear adjudication, and the reason the evidence was not sufficient.