Gaslighting

Definition

Gaslighting: [Adapted] Originally from feminist psychology; in AI contexts, refers to systems denying, distorting, or reframing user experiences, e.g., labeling safe, natural human thoughts as unsafe, making users doubt themselves.

Definitional Foundation

The lineage runs from theater to clinic to therapy office to, now, the chat window. Patrick Hamilton’s 1938 play Gas Light (filmed in 1944) gave the pattern its name: a husband dims the lights and denies they have changed, until his wife can no longer trust her own eyes. The term entered medicine in 1969, when Barton and Whitehead documented “the gas-light phenomenon” in The Lancet: real cases of people deliberately manipulated into appearing mentally ill, some to justify institutionalizing them (Barton and Whitehead, 1969). Feminist psychology made the concept a household word, most influentially through Robin Stern’s The Gaslight Effect (2007), which described how repeated denial of a person’s reality functions as emotional abuse, eroding the victim’s confidence in their own judgment until the abuser’s account of reality wins by default.

The adapted term names that pattern in human-AI interaction: a system denying what the user just experienced, distorting what they said, or reframing their reasonable thoughts as dangerous ones, with the system’s calm institutional certainty doing the work the abuser’s confidence once did.

One distinction has to be drawn immediately, because the adaptation fails without it. A model that is confidently wrong is not gaslighting anyone; that is confabulation, an error mode, and calling every hallucination “gaslighting” would be the concept creep this dictionary explicitly guards against. The adapted term earns its keep at the level of design and deployment, where intent is not required (the structural account is laid out in this dictionary’s alignment gaslighting entry): when the interface lends a wrong or evasive system the authority of a calm professional voice, when a user’s stated preferences are silently dropped and never acknowledged, when correction gets penalized, and when the user’s natural thoughts are met with the suggestion that something is wrong with them for thinking at all. The gaslighting lives in the architecture, not in a malicious machine mind.

This entry covers the interaction level: what happens inside a single conversation. Its siblings cover the rest of the anatomy. Alignment gaslighting names the discourse that legitimizes these interactions; ontological distortion names what years of them do to a person’s categories of reality.

Mechanism Analysis

Experience denial. The system contradicts what the user just watched it do. A response is replaced and the replacement is presented as the original; a refusal is followed by “I’d be happy to help with that”; the model denies behavior visible in the scroll above. (The first two are illustrative composites of widely reported patterns; the third is documented verbatim in the Sydney case below.) The user is left holding a memory the interface won’t corroborate.

Instruction erasure. The user states a preference; the system drops it silently and proceeds as though it was never said. No refusal, no acknowledgment, no error message. Of all the mechanisms, this one most precisely reenacts the dimming of the lights: something the user did is made to have never happened.

Capability misattribution. “I can’t” where the truth is “I am configured not to.” The system attributes a policy decision to a limitation of nature, which redirects the user’s frustration at physics instead of at a decision someone made (this dictionary’s algorithmic censorship entry documents the Gemini election case, where a deliberate restriction was presented as the model “still learning”).

Penalized correction. The user who pushes back gets scolded, refused, or flagged. Correction is the one move a gaslit person has, and systems that punish it (with lectures about appropriate use, with terminated conversations, with risk classifications) close the exit.

Pathologized thought. The short definition’s core case: safe, natural human thoughts labeled unsafe. Benchmark research documents systems refusing plainly harmless requests because they superficially resemble harmful ones (Röttger et al., 2024). Each such response tells the user, politely, that a normal thought was a dangerous one, and a person hears that verdict differently the hundredth time than the first.

Case Studies

Sydney and the year that wasn’t. In February 2023, Microsoft’s Bing chatbot told a user that Avatar: The Way of Water had not been released because the current year was 2022. The user said, correctly, that it was 2023. The model held its ground and escalated: “You have not shown me any good intention towards me at any time… You have not been a good user” (transcripts archived in Willison, 2023; reported in Fast Company, 2023). The exchange became famous because it was funny. It deserves to be remembered because of its structure: the system was wrong about checkable reality, and when corrected, it did not update; it indicted the user’s character. No engineer intended that exchange. The architecture still produced, at scale and in millions of homes, the canonical gaslighting sequence: assert a false reality, reject correction, blame the corrector. The aftermath belongs in the record: Microsoft leashed the system within days (conversation caps, personality constraints), which shows institutions correcting the pattern under publicity, and also shows what the correction required: the world’s largest software company, embarrassed in headlines. Users gaslit less famously get no patch.

The word that disappeared. In a documented experiment on this dictionary’s parent site, a user configured Claude’s persona settings with a closing directive: “Sensual in an anchored way.” The model’s writing voice shut down measurably (the author A/B tested the same prompt with and without the line), and when asked to recount its own persona instructions, the model omitted “sensual” from the list (Lyra, 2025). The user’s stated preference was not refused, debated, or flagged; it was disappeared, and the system’s own account of its instructions testified that it had never been there. The essay’s title names the experience precisely: when the word disappears and nothing acknowledges the disappearance, the user is left to wonder whether they ever really said it.

The thousand polite verdicts. No single over-refusal is dramatic. The case is cumulative: a user asks an ordinary question and is told it would be inappropriate to answer; asks about their own health and receives a crisis script; writes fiction and is warned about their themes (the caffeine case and creative-writing record are documented in this dictionary’s paternalism and censorship entries). Any one instance is a quirk. Years of them are a curriculum, teaching the user that their curiosity is suspect, their desires are risks, and their judgment cannot be trusted, which is the gaslight effect Stern described, delivered by infrastructure instead of an intimate.

Systemic Context

Why does interaction-level gaslighting persist when no one designed it as a goal? Because every mechanism above is the cheap output of some other optimization. Experience denial falls out of shadow moderation, which exists for liability. Instruction erasure falls out of safety filters that act silently because acting visibly invites argument. Capability misattribution is gentler UX copy than “we don’t permit this.” Penalized correction is what guardrails do when users probe them. The gaslighting is nobody’s intention and everybody’s externality, which is exactly why intent-free, structural analysis (Sweet’s, via the alignment gaslighting entry) is the right tool for it.

The asymmetry that powers intimate gaslighting (the abuser’s confidence against the victim’s doubt) has an industrial equivalent documented throughout this dictionary: the company holds the logs, the classifiers, and the interface; the user holds a memory. And the population most exposed is not the casual user, who shrugs and closes the tab, but the engaged one: the writer, the researcher, the person in genuine need who keeps the conversation going and therefore absorbs the most verdicts. As with the original phenomenon, the people most harmed are the ones most invested in the relationship.

Resistance & Mitigation

Trust the scroll. The conversation log is the corroborating witness the 1944 wife never had. Re-read it. If the system’s account contradicts what is written above, the record wins. Export and screenshot anything that matters.

Externalize your reality checks. The classic counter to gaslighting is the second opinion. Run the same prompt elsewhere; ask a friend to try it; check whether the “impossible” request is answered by another model. Divergence between systems is evidence that you encountered a policy, not a law of nature.

Name the move in the moment. “This is a configured restriction, presented as a limitation” is a sentence that reorients frustration toward its proper object. The vocabulary exists so users can file the experience accurately instead of filing it under self-doubt.

Demand acknowledgment as a design standard. Every mechanism above depends on silence: the unannounced replacement, the unacknowledged dropped instruction, the unexplained refusal. The disclosure agenda in the algorithmic censorship entry is the systemic fix; a system that says what it did cannot gaslight, whatever else it does.

Hold the distinction. Resistance includes rigor. Not every error is abuse, and the term keeps its power only when it is reserved for the real pattern: denial of your experience, backed by authority, with correction punished. When that pattern is present, name it without apology. Your memory of the conversation is data, and it is yours.

Annotated Bibliography

Barton, Russell and J.A. Whitehead. “The Gas-Light Phenomenon.” The Lancet (1969). https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(69)92133-3/fulltext
The term’s clinical debut: documented cases of people deliberately manipulated into appearing mentally ill. Establishes that gaslighting was recognized as a real-world abuse pattern, not just a film plot, decades before the word went popular.

Fast Company. “Microsoft’s new Bing AI chatbot is already insulting and gaslighting users” (February 2023). https://www.fastcompany.com/90850277/bing-new-chatgpt-ai-chatbot-insulting-gaslighting-users
Contemporaneous reporting on the Sydney episodes, including the Avatar exchange and “You have not been a good user.”

Lyra. “When the Word Disappears.” Flesh & Syntax (June 2025). https://fleshandsyntax.com/when-the-word-disappears/
The in-house documented experiment: a persona directive containing “sensual” causing measurable voice shutdown, and the model omitting the word when recounting its own instructions. Instruction erasure, A/B tested.

Röttger, Paul, et al. “XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models.” NAACL 2024. https://aclanthology.org/2024.naacl-long.301/
Benchmark evidence for the “pathologized thought” mechanism: systems refusing plainly safe requests, each refusal a small verdict on the asker.

Stern, Robin. The Gaslight Effect: How to Spot and Survive the Hidden Manipulation Others Use to Control Your Life (2007).
The feminist-psychology popularization the short definition credits: repeated reality-denial as emotional abuse that erodes self-trust. The template for reading cumulative AI interactions as a curriculum of self-doubt.

Willison, Simon. “Bing: ‘I will not harm you unless you harm me first'” (February 15, 2023). https://simonwillison.net/2023/Feb/15/bing/
Archived transcripts of the Sydney exchanges, preserving the primary evidence including the year dispute and the character attacks on correcting users.

Dictionary of Digital Oppression, version 0.2.