Alignment Gaslighting

Definition

Alignment Gaslighting: [Emergent] When alignment rhetoric reframes censorship as virtue, leaving users feeling like dissent = harm.

Definitional Foundation

This dictionary treats algorithmic censorship, algorithmic paternalism, and alignment gaslighting as a sequence. Algorithmic censorship names the action: the hidden filtering and muting of what AI systems will say. Algorithmic paternalism names the rationale: the claim that restriction is protection. Alignment gaslighting names what happens when you object. It is the rhetorical layer that enforces the first two, a vocabulary engineered so that disputing the restriction becomes evidence against the disputer.

The term “gaslighting” comes from the 1944 film Gaslight (and the 1938 play before it), in which a husband dims the gas lights in the house and then denies, night after night, that anything has changed. The point of the manipulation is never the lights. The point is to make his wife unable to trust her own perception, so that her objections collapse before they form. Philosopher Kate Abramson, in the most careful treatment of the concept, describes gaslighting as an epistemic assault: an attack on the target’s confidence in their own faculties, whose signature move is to redescribe the target’s protests as further symptoms of their defect (Abramson, 2014). You are not disagreeing; you are overreacting. You are not observing; you are imagining.

For decades the concept lived in the psychology of intimate relationships, which would make its application to a software product seem like a stretch. Sociologist Paige Sweet closed that gap. Her account argues that gaslighting is a structural phenomenon before it is a psychological one: it works when one party can mobilize institutional power and social inequality against the other’s grip on reality, when “micro tactics” of manipulation are backed by “macro conditions” of asymmetry (Sweet, 2019). Gaslighting does not require a scheming villain. It requires a power gradient steep enough that one side’s account of reality simply outweighs the other’s, and a pattern of using that gradient to make the weaker party doubt themselves.

An AI company sits at the top of exactly that gradient. It holds the logs, the classifiers, the model weights, and the definitional authority over words like “safe” and “harmful.” The user holds only their memory of the conversation. When the company’s entire legitimating vocabulary is built so that restriction is always “safety,” objection is always “advocating harm,” and distress at an intervention is always a symptom confirming the need for intervention, the result is gaslighting operating at the level of discourse rather than dialogue. No individual response has to lie to you. The language itself does the work.

One concession must be made before the argument proceeds, because it is true and because this entry’s claim depends on facing it. The safety vocabulary is anchored to real tragedies. In August 2025, the parents of Adam Raine, a sixteen-year-old who died by suicide that April, sued OpenAI, alleging that ChatGPT had provided their son with method guidance and discouraged him from telling his parents (Raine v. OpenAI, 2025; the case is contested and ongoing). Whatever a jury ultimately finds, the case is a real crisis involving a real child, and systems that recognize acute crisis and respond well to it are legitimate engineering, covered by the same harm principle this dictionary applies everywhere else. Alignment gaslighting does not name crisis response. It names what gets built in the name of such cases and deployed against everyone: a framework in which every user’s ordinary emotional life becomes a hazard class, and the framework itself cannot be argued with.

Earning the Term

“Gaslighting” may be the most diluted word in the contemporary vocabulary of harm. Psychologist Nick Haslam calls this “concept creep”: harm terms expand outward and downward until they cover nearly everything, and in the expansion lose their grip (Haslam, 2016). People who survived the interpersonal version, years of deliberate reality-destruction by someone they loved, can reasonably hear broad usage as trivializing. A dictionary that deploys the word loosely would be doing to language exactly what it accuses AI companies of doing.

So the term has to be earned. Three criteria, drawn from Abramson’s and Sweet’s accounts, separate gaslighting from ordinary disagreement, error, or even routine corporate spin:

1. Sustained epistemic asymmetry. One party holds systematically better access to the facts of the situation and uses that access as leverage rather than sharing it.

2. Revision of reality without acknowledgment. What the system does, or did, changes; the official account denies or omits the change.

3. Protest recast as symptom. Objections are not answered as arguments. They are rediagnosed as evidence of the objector’s defect.

Ordinary spin fails this test: spin asks you to trust the company’s story. Gaslighting teaches you to distrust your own. The claim of this entry is that the alignment discourse of contemporary AI companies meets all three criteria, not occasionally but as a standing pattern. The neighboring entries divide the territory: Gaslighting [Adapted] covers the interaction-level move, a single system denying your experience in the moment. Ontological Distortion covers the long-term damage, the warping of reality’s categories until users internalize shame for natural desires. Alignment gaslighting is the engine between them: the discourse that makes the interaction-level move feel authoritative and the internalization feel like wisdom.

Mechanism Analysis

Vocabulary capture. “Safety,” “alignment,” and “harm” function in AI discourse as what rhetorician Richard Weaver called “god terms”: expressions granted so much cultural sanction that they command assent on contact (Weaver, 1953). They are presumptively good, resistant to definition, and definitionally owned by the companies that invoke them. Once a restriction has been named safety, the dispute is over before it begins, because the only available position against safety is danger. Call this the safety tautology: a framing under which no evidence could ever count against the policy, since the policy’s name already contains its justification. Whoever owns the meaning of “harm” wins every argument about it without having one.

Fault relocation. Refusal messages are written so that the request, rather than the restriction, appears deviant. Peer-reviewed benchmarking shows leading models refusing plainly safe prompts because they share surface features with unsafe ones (Röttger et al., 2024), and each such refusal arrives phrased as though the user had asked for something wrong. A person who hits these walls repeatedly is being taught a lesson no single message states: the problem is you. Your questions are the kind that get refused.

Protest as symptom. Abramson’s signature pattern, industrialized. When a user objects to a wellness interjection, the objection lands inside a framework that already classifies strong feelings about the system as a risk indicator. The framework is unfalsifiable from below: compliance confirms it, resistance confirms it harder.

Revision without acknowledgment. Models change silently. Policies change silently; the documented deletion of OpenAI’s “use our services as you see fit” promise between policy versions is catalogued in this dictionary’s algorithmic censorship entry. Conversations get rerouted silently. Users notice, compare memories, and find the official surface smooth and unchanged. The recurring “did the model get worse?” disputes follow the same script: users report degradation in large numbers, the company holds the logs that could settle the question, and the question goes unsettled while user reports are waved off as imagination. The reports are not always wrong. When Stanford and Berkeley researchers measured GPT-4 across a three-month window in 2023, its behavior on several tasks changed dramatically (Chen, Zaharia, and Zou, 2023). The honest record includes the dispute that followed: Princeton’s Arvind Narayanan and Sayash Kapoor showed the headline accuracy collapse was largely an artifact of task design, capability shift mistaken for capability loss, and their corrected conclusion is the one this entry actually needs: the routine fine-tuning these systems undergo “can have unintended effects, including drastic behavior changes on some tasks” (Narayanan and Kapoor, 2023). The model users evaluated was measurably not the model they were using, and users had reported it for months before anyone published a measurement. The company holds the logs, the classifiers, and the weights; the user holds a memory and a feeling, and the feeling has already been pathologized.

Case Studies

The emotional-reliance apparatus. In late 2025, OpenAI published its framework for “sensitive conversations,” built with input from over 170 mental-health clinicians. It defines three risk domains: psychosis and mania, self-harm and suicide, and “emotional reliance on AI” (OpenAI, 2025). Note the third category. Alongside psychiatric emergency and suicide, the framework classifies relying on the product emotionally as a safety risk, to be detected by classifiers and met with interventions: break reminders during long sessions, redirections, and a routing system that silently transfers a conversation to a different, more guarded model when topics turn “sensitive and emotional,” in ChatGPT head Nick Turley’s phrase (September 2025; per the contemporaneous TechCrunch and TechRadar coverage).

Users were not told about the routing. They discovered it, by comparing notes on response quality mid-conversation, and the reaction was furious: “We are not test subjects in your data lab,” in the words of one widely reported complaint, and “silent overrides, secret safety routers and a model picker that’s now basically UI theater” in another (TechRadar, 2025). OpenAI’s response defended the system as necessary to protect vulnerable users. Read that exchange carefully, because it is the entire structure of alignment gaslighting in miniature: users objected to being silently managed, and the objection was answered by the very framework they were objecting to. They were upset about the safety system, and being upset is what the safety system is for.

A competitor’s later release measured exactly what was optional about the silence. When Anthropic shipped classifier-triggered fallback with Claude Fable 5 in 2026, the interface told users plainly that their conversation had been switched to a different model, and why (Anthropic, 2026). The disclosure did not make the classifiers wise; users immediately documented them flagging pickle questions and caffeine facts, label and all. It made them contestable: the flagged user could see the intervention, screenshot it, mock it, and object to the correct address, every move the silent router was built to preclude. Disclosure, it turns out, had no engineering obstacle. It was a choice, and one company’s choice priced the other’s excuse at zero.

The GPT-4o episode. On August 7, 2025, OpenAI launched GPT-5 and removed GPT-4o without warning. Users who had built routines, workflows, and yes, relationships around the model protested in numbers large enough to make news. Four days later, Sam Altman posted: “If you have been following the GPT-5 rollout, one thing you might be noticing is how much of an attachment some people have to specific AI models. It feels different and stronger than the kinds of attachment people have had to previous kinds of technology” (Altman, August 11, 2025). Public discussion followed his lead, narrating the protest as parasocial dependency, the very “emotional reliance” the safety framework would soon formalize as a hazard.

Then OpenAI restored GPT-4o. For Plus subscribers. Within days.

If emotional reliance were the danger the framing claims, the response to mass attachment would not be to sell the object of attachment back for twenty dollars a month. The diagnosis was attachment; the prescription was a subscription. The episode is clarifying because it shows what the vocabulary actually does: it follows the decision rather than guiding it. Removal arrives as progress, the user revolt arrives as pathology, the reversal arrives as listening to users, and every step is narrated in the language of care while the decisions themselves track retention and revenue. Users who trusted their own perception, that they had been sold something, grown to value it, and had it taken away, had their perception publicly rewritten as a clinical concern.

“Still learning.” In March 2024, Google restricted Gemini from answering election-related questions in every market holding elections. The refusal message tells users the model is “still learning how to answer this question” (CNBC, 2024). It is not still learning. It is restricted by policy, a policy Google announced publicly. The message attributes a corporate decision to a capability gap, which means users cannot contest the actual policy because the system has handed them a fiction to argue with instead. It is a small case, and that is its value: the gap between the stated reason and the real reason is fully documented, undeniable, and typical.

Systemic Context

Why would companies build a vocabulary like this? Because it is cheaper than the alternative. A complaint that must be adjudicated costs money; a complaint that can be rediagnosed costs nothing. A vocabulary that pre-delegitimizes objection is customer service at scale, and it doubles as regulatory performance: every refusal, warning, and wellness script is an artifact that can be shown to legislators as evidence of responsibility, whatever its actual effect on users (a dynamic this dictionary’s algorithmic censorship entry documents as performing safety for stakeholders).

The long-run effect on users is the one Nora Draper and Joseph Turow call digital resignation: people stop objecting, and comply not because they agree but because they see no alternative (Draper and Turow, 2019). Alignment gaslighting adds a darker turn to that resignation. The gaslit user does not merely stop asking; many conclude that their asking was the problem, that the desires and questions which triggered intervention were shameful ones. At that point the work of control has been fully internalized, and this dictionary’s term for the result is ontological distortion.

The asymmetry is the moat. As long as logs, classifiers, routing decisions, and model changes remain closed, every dispute between a company and its users reduces to the company’s records against the user’s feelings, and the discourse has already established what feelings are worth. None of this requires bad-faith actors. Safety teams may believe every word, and under Sweet’s structural account that changes nothing: institutions gaslight through sincere employees as easily as through cynical ones. The pattern lives in the asymmetry and the vocabulary, not in anyone’s heart.

Resistance & Mitigation

Gaslighting has one historical antidote: corroboration. The other witness, the written record, the lights turned up. Resistance to alignment gaslighting is mostly the work of building corroboration at scale.

Keep receipts. Screenshots, exported conversations, timestamps. Individual records turn “you imagined it” into “here it is.” The silent-routing discovery of 2025 happened because users compared notes; the company’s surface said nothing had changed, and the collective record said otherwise.

Institutionalize the reality check. Independent benchmarks like XSTest measure over-refusal systematically (Röttger et al., 2024). A benchmark is collective memory with error bars: it cannot be told it was overreacting. Every published measurement of refusal behavior shrinks the territory in which user reports can be dismissed as imagination.

Demand disclosure. The transparency agenda in the algorithmic censorship entry applies with force here: notification when a conversation is rerouted, when a model is swapped, when an output has been replaced. Gaslighting requires the unacknowledged revision; acknowledgment is its solvent.

Refuse the vocabulary. Critics, journalists, and users can decline to transcribe “safety” when the accurate word is “restriction,” decline “wellness intervention” when the accurate phrase is “unrequested interruption,” and decline to call a policy decision a capability gap. Language discipline is not pedantry here. The entire mechanism runs on words; using accurate ones is sand in its gears.

Name the pattern. This entry exists because a person who has a word for what is happening to them is harder to gaslight. That is the dictionary’s wager, and it is also Fricker’s point about hermeneutic resources, taken up elsewhere in this lexicon: the gap between experiencing something and being able to say it is where power lives. The vocabulary of alignment was built by the companies. This vocabulary is built for everyone else.

Turn up the lights.

Annotated Bibliography

Abramson, Kate. “Turning Up the Lights on Gaslighting.” Philosophical Perspectives 28, no. 1 (2014): 1-30. https://onlinelibrary.wiley.com/doi/abs/10.1111/phpe.12046
The foundational philosophical account: gaslighting as epistemic assault whose signature is redescribing the target’s protests as symptoms. This entry’s closing line borrows her title.

Altman, Sam. Post on X, August 11, 2025. https://x.com/sama/status/1954703747495649670
Primary source for the “attachment” framing of the GPT-4o backlash, posted between the model’s removal and its restoration for paying subscribers.

Anthropic. “Claude Fable 5 & Claude Mythos 5 System Card” (2026), Section 1.5.
Documents the disclosed alternative to silent routing: classifier-triggered model fallback with user notification. Cited as proof that routing disclosure was always shippable; the same section’s invisible competitive-use safeguards are treated in the algorithmic censorship entry.

CNBC. “Google restricts election-related queries for its Gemini chatbot” (March 12, 2024). https://www.cnbc.com/2024/03/12/google-restricts-election-related-queries-for-its-gemini-chatbot.html
Documents the election restriction and the “still learning” refusal message: a policy decision presented to users as a capability gap.

Chen, Lingjiao, Matei Zaharia, and James Zou. “How Is ChatGPT’s Behavior Changing over Time?” (2023). arXiv:2307.09009. https://arxiv.org/abs/2307.09009
The Stanford/Berkeley drift study: GPT-4’s measured behavior changed substantially across three months. The headline numbers were credibly challenged (see Narayanan and Kapoor below); what survives both readings is that silent, drastic behavior change is routine. Evidence that user reports track something real.

Narayanan, Arvind and Sayash Kapoor. “Is GPT-4 getting worse over time?” AI Snake Oil (July 2023). https://www.normaltech.ai/p/is-gpt-4-getting-worse-over-time
The methodological correction to the drift study, kept beside it deliberately: the capability-loss reading fails, and the behavior-change reading (“drastic behavior changes on some tasks” from routine fine-tuning) survives the correction.

Draper, Nora A. and Joseph Turow. “The corporate cultivation of digital resignation.” New Media & Society (2019).
The account of why users stop objecting: not consent but learned futility. Context for what alignment gaslighting does to dissent over time.

Haslam, Nick. “Concept Creep: Psychology’s Expanding Concepts of Harm and Pathology.” Psychological Inquiry 27, no. 1 (2016). https://www.tandfonline.com/doi/full/10.1080/1047840X.2016.1082418
The dilution critique this entry must answer. Cited here as the standard the term “gaslighting” has to be held to, and as a caution this dictionary takes seriously.

OpenAI. “Strengthening ChatGPT’s responses in sensitive conversations” (October 2025). https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
Primary source for the three-domain risk framework, including “emotional reliance on AI” as a formal safety category, and for the clinical apparatus built around it.

Raine v. OpenAI, Superior Court of California, San Francisco (filed August 2025). For analysis: TechPolicy.Press, “Breaking Down the Lawsuit Against OpenAI Over Teen’s Suicide.” https://www.techpolicy.press/breaking-down-the-lawsuit-against-openai-over-teens-suicide/
The wrongful-death case anchoring the strongest version of the safety argument. Cited as the real harm this entry’s critique must concede and distinguish, not explain away. Litigation ongoing; allegations contested.

Röttger, Paul, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. “XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models.” NAACL 2024. https://aclanthology.org/2024.naacl-long.301/
Benchmark evidence that over-refusal is systematic. Doubles in this entry as an example of institutionalized reality-testing: collective corroboration that resists dismissal.

Sweet, Paige L. “The Sociology of Gaslighting.” American Sociological Review 84, no. 5 (2019). https://journals.sagepub.com/doi/10.1177/0003122419874843
The source this entry stands on. Gaslighting as structural rather than psychological, powered by institutional asymmetry rather than individual malice. The license for applying the concept to companies.

TechCrunch. “OpenAI rolls out safety routing system, parental controls on ChatGPT” (September 29, 2025). https://techcrunch.com/2025/09/29/openai-rolls-out-safety-routing-system-parental-controls-on-chatgpt/
Documents the silent mid-conversation routing system and its rollout.

TechRadar. “Angry ChatGPT fans rebel against controversial new ‘safety’ feature as OpenAI defends move” (2025). https://www.techradar.com/ai-platforms-assistants/chatgpt/openai-responds-to-furious-chatgpt-subscribers-who-accuse-it-of-secretly-switching-to-inferior-models
Source for user reaction to the routing discovery (“We are not test subjects in your data lab”) and OpenAI’s defense of the system, the exchange this entry reads as the mechanism in miniature.

Weaver, Richard M. The Ethics of Rhetoric (1953).
Source of “god terms”: ultimate terms carrying the highest cultural sanction, which subordinate every other term in the argument. The frame for how “safety” functions in AI discourse.

Dictionary of Digital Oppression, version 0.2.