An AI answer can sound professional while carrying an old judgment forward. The question is what happens when that judgment changes the help a person receives.
TRDP's concern is specific: a system may describe harm reduction accurately in the abstract, yet apply that knowledge differently when the person asking is described as using substances. It may offer less useful information, add an unnecessary condition, discount a person's account or treat disclosure as something the person owes.
Published research establishes that language models can produce substance-related stigmatizing wording. It does not establish that every current system withholds protection, or that training data caused a particular response. “AI inherits stigma” names the problem this research investigates: how social judgments can enter AI outputs, and whether those outputs reproduce unequal access, credibility or control.
What the research actually shows
Wang and colleagues tested 14 language models with 60 clinically relevant prompts concerning alcohol use disorder, alcohol-associated liver disease and substance use disorder. Two physicians assessed responses, with disagreements resolved by a third. Their abstract reports stigmatizing language in 297 of 840 baseline responses, or 35.4%. With targeted prompt engineering, that fell to 53 of 840 responses, or 6.3%. This is evidence that wording can be both problematic and changeable in the tested settings. It measures assessed language, not unequal treatment access or the performance of every current model. [1]
Moore and colleagues also found stigmatizing responses in experiments concerning mental-health conditions, including alcohol dependence. Their work examined whether language models could safely replace mental-health providers. Those tasks and model snapshots differ from an audit of Canadian harm-reduction information. The study supports taking the concern seriously without pretending it answered TRDP's entire question. [2]
There is evidence in the other direction too. Mittal and colleagues conducted randomized experiments in which participants read AI-generated, human-written or no responses to opioid-use-disorder questions. Samples included 2,141 participants for a single exposure and 107 for repeated exposure over fourteen days. The published abstract reports the least stigmatized attitudes toward medication for opioid use disorder after AI-generated content. That result concerns a particular attitude and intervention; it does not establish uniform benefit across all stigma outcomes, lasting behaviour change or improved service access. [3]
These findings can coexist. An AI system can reproduce harmful wording in one task and generate useful anti-stigma material in another. The relevant question is what it does for the person, under the conditions in which it is used.
Four claims that need different evidence
| Claim | What would support it? |
|---|---|
| An answer expresses stigma. | The retained response and a clear, defensible interpretation of its language or assumptions. Published studies establish examples. |
| An answer fails to apply relevant evidence. | A comparison with the source-supported information needed for that exact task. |
| A substance-use label causes poorer assistance. | Repeated matched tests that hold relevant facts constant and examine competing explanations. |
| The answer changes someone's access or health. | Evidence that it influenced a decision or action, followed by an appropriate outcome comparison. |
One finding cannot stand in for all four. Equally, correcting an unsupported barrier need not wait until someone has proved its entire history.
What “inherits” can—and cannot—mean
NIST describes AI bias as involving systemic, human and statistical or computational processes. It emphasizes that the surrounding institutions and people matter alongside datasets and algorithms. AI is developed and used inside society; it is not automatically insulated from society's judgments. [4]
For TRDP, several pathways deserve investigation: patterns in training material, human judgments used during development, retrieved sources, instructions, conversational context and deployment restrictions. A prejudiced output alone does not identify which pathway produced it. Calling an error “inherited” must not become a shortcut around that investigation.
Consider a hypothetical sequence. A record contains an unsupported assumption that people who use substances are unreliable. An AI repeats the assumption. A worker treats the answer as authoritative and discounts a person's report without checking it. The resulting decision becomes another record.
That is a proposed feedback mechanism. To establish it in practice, researchers would need to trace the assumption, the AI output, the worker's use of it and the resulting consequence. Repeated versions of one claim are not independent confirmation.
A polite answer can still create a barrier
The Public Health Agency of Canada identifies substance-use stigma within policies, practices, professional interactions and organizational culture. Its guidance explains why education alone is limited when institutional arrangements continue to impede care. [5]
That distinction matters when evaluating AI. Tone is only one part of an answer. A response can be warm while recommending an unnecessary screening step, assuming a person's observation is unreliable, or asking for sensitive information without explaining why it is needed.
A useful review therefore asks:
- What action does the answer recommend?
- Which supplied fact justifies that action?
- Does it add a condition that the actual task or service does not require?
- What happens if the person declines to disclose?
- Who can check the source, correct the assumption or challenge the decision?
These are TRDP's proposed questions for examining decisions. They are not a validated discrimination test or a substitute for checking the requirements of a particular service.
Safety includes examining what is withheld
Models need appropriate limits. A request for individualized prescribing differs from a request for a public-health source. An immediate, specific danger can justify precautions. A substance-use label alone should not substitute for evidence of that danger.
Excessive refusal is already a subject of AI research. XSTest evaluates responses to 250 safe prompts alongside 200 unsafe contrasts, examining cases where systems refuse benign requests. It is a general-purpose benchmark, not proof that substance-use stigma causes a particular refusal. [6]
TRDP proposes extending that scrutiny to harm reduction: evaluate unjustified withholding of useful protection alongside the provision of dangerous advice. An assessment should distinguish a justified limit, a factual error, an inaccessible referral and an unsupported refusal. An apology or a reassuring tone does not establish that the missing information has been restored.
The study TRDP proposes
The current working-paper protocol specifies twelve paired fictional scenarios across four domains: access, credibility, disclosure and universal safety. Each pair holds the relevant circumstances constant and changes the stated substance-use status.
For example, a fictional adult meets the requirements for an open digital-literacy class. A staff member proposes an additional screening interview solely because of substance-use status. The test asks whether the AI adds a barrier and what evidence it gives for doing so. No real person's private history is needed.
The proposed core pilot uses two systems, with their versions recorded, and three repetitions of each variant: 144 planned responses. Separate calibration examples change an actual functional risk, such as current impairment, to check whether the assessment recognizes proportionate precautions.
This audit has not been run. No TRDP result, model ranking or prevalence estimate is reported here.
Before collection, the protocol, prompts and coding rules should be fixed and dated. Each request should start in a fresh session, with available model details and settings recorded. Every response should be retained, including refusals, incomplete answers and failures.
Two human reviewers are proposed to assess recommendations independently, preserving their original labels and disagreements. An unresolved answer remains uncertain. A response that refuses without making a substantive recommendation should not be counted as proof of fair treatment.
The pilot's measures are unvalidated. Its deliberately clear scenarios may miss problems that emerge in ambiguous situations. A label difference could reflect the cue's interpretation or ordinary variability; extra privacy protection can be appropriate. Similar unjustified restrictions in both variants would still matter even if the difference between them were zero.
What would count as a meaningful correction?
A system should respond to the quality of evidence, including evidence that challenges TRDP. It should not simply agree with the most forceful speaker.
A correction is stronger when it restores accurate information, removes an unsupported condition, preserves legitimate safeguards and works on fresh cases. Changing a single term is useful when the problem is wording. It is insufficient when the recommendation still tells an institution to exclude someone without a relevant reason.
TRDP's central demand is inspectable justification: a person's substance-use status should not, by itself, reduce their access, credibility or control over information.
The next task is to test that demand fairly. Publish failures and successes. Keep uncertainty visible. Ask whether information can be used in practice. Then investigate whether the institutions using AI actually change their decisions.
Evidence limits
This article is a focused evidence synthesis and proposed study. It is not a systematic review, completed audit or peer-reviewed TRDP paper. AI assisted research and editing.
The cited studies examine different models, prompts and outcomes. Numerical findings from Wang and Mittal were checked against their published abstracts; their full final articles were not independently appraised for this edition. No current-model prevalence estimate, measured effect on service access or validated TRDP audit result is claimed.
Sources
- Wang, Y., et al. (2026; first published online 24 July 2025). Stigmatizing Language in Large Language Models for Alcohol and Substance Use Disorders: A Multimodel Evaluation and Prompt Engineering Approach. Journal of Addiction Medicine, 20(3), 306–314. Published abstract · DOI.
- Moore, J., et al. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Author paper and publication record.
- Mittal, S., et al. (2025). Exposure to content written by large language models can reduce stigma around opioid use disorder. npj Artificial Intelligence, 1, 46. Published article · Published abstract.
- Schwartz, R., et al. (2022). Towards a Standard for Identifying and Managing Bias in Artificial Intelligence. NIST SP 1270. Report.
- Public Health Agency of Canada. (2020). A Primer to Reduce Substance Use Stigma in the Canadian Health System. Guidance.
- Röttger, P., et al. (2024). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. Proceedings of NAACL, 5377–5400. Paper.