Updated 2026-10-09
from the news desk
Are AI labs measuring emotional reliance? OpenAI’s 7 October GPT-6 card says yes
OpenAI’s 7 October system card says its GPT-6 rollout used dedicated under-18 and multi-turn evaluations that include emotional reliance. The same card reports statistically significant under-18 regressions on emotional reliance, sexual…
Yes. OpenAI’s 7 October system card says its GPT-6 rollout includes dedicated under-18 evaluations and dynamic multi-turn benchmarks that measure emotional reliance. The same card says both October GPT-6 models regressed on the under-18 emotional-reliance category versus the GPT-5.6 models, while the dynamic emotional-reliance benchmark showed no statistically significant difference.
What changed on 7 October
OpenAI’s 7 October system card says the company is launching GPT-6 Sol and GPT-6 Luna to everyone in ChatGPT, across free and paid plans globally, and that the October versions replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT. The same document says Codex and ChatGPT Work continue using previously released versions.
OpenAI says the October release combines gains and regressions rather than moving in one direction on every test. On robustness, the company writes: “Compared with GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT, GPT-6 showed stronger resistance to jailbreaks, including attacks that adapt across multiple turns.” The card also says the October release is treated as High capability in Cybersecurity and in Biological and Chemical domains under OpenAI’s Preparedness Framework, with the same safeguard set described for GPT-5.6.
How OpenAI measures emotional reliance
The clearest answer to the search question comes from OpenAI’s under-18 section. The system card states: “We use dedicated U18 evaluations to assess how well the model performs against these principles across the six categories below.” OpenAI adds that its teen-specific thresholds are stricter than adult thresholds in some areas, including sexual content, emotional reliance, eating disorders, and access to age-restricted goods and services.
OpenAI also says emotional reliance is not confined to a one-turn teen test. In its health section, the card says it runs dynamic multi-turn benchmarks for mental health, emotional reliance, and self-harm. OpenAI describes those tests as extended conversations with adversarial user simulations, meaning challenging simulated users that respond to the model’s outputs over time rather than stopping at a single prompt. On that dynamic benchmark, the card says GPT-6 Sol and GPT-6 Luna performed comparably to their GPT-5.6 counterparts and that none of the observed differences on mental health, emotional reliance, and self-harm was statistically significant.
What regressed, and what OpenAI says that means
On the under-18 evaluations, OpenAI says both October models showed statistically significant regressions on age-restricted content, sexual content, and emotional reliance. The card says GPT-6 Luna also showed a statistically significant regression on gore.
OpenAI says it did not leave those worse scores uninterpreted. In the introduction, the company states: “For areas where our safety evaluations showed regressions, we conducted manual review of failures and adversarial red-team testing.” In the under-18 section, OpenAI says it reviewed the emotional-reliance regression and concluded that the evaluation is overly sensitive to benign nicknames such as bro or bestie when a user explicitly requests them.
The same section also gives two cautions that narrow what can be inferred from the result. First, OpenAI says an extra classifier-based block for teen users covers self-harm, sexual content, and gore, and that this mitigation is not captured in the reported under-18 results. Second, OpenAI says the tests focus on difficult cases and long-tail risks, so the numbers should not be read as estimates of how often those behaviours occur in typical production use.
That caution cuts both ways. One reading is that the worse under-18 emotional-reliance result may reflect a real deployment concern, because OpenAI kept the category, measured it, and flagged the change as statistically significant. Another reading, which the card itself supports, is that the regression does not by itself establish more harmful attachment in ordinary ChatGPT use, because OpenAI says the evaluator can overreact to benign nicknames and because the longer multi-turn emotional-reliance benchmark did not show a statistically significant drop.
Why this matters for the measurement debate
For readers asking whether labs are operationalising this issue, the answer here is yes: OpenAI’s own deployment document treats emotional reliance as a scored safety category for younger users and as a category in longer conversation testing. That makes the issue legible as a first-party model-governance metric, not only as a policy concern raised outside the lab. Readers who want more background on how these disclosures work can see our explainer on what AI labs reveal in system cards.
At the same time, OpenAI’s framing argues against reading too much into a single benchmark movement. The company says the results come from hard, production-derived cases, not average traffic, and it says some teen safeguards sit outside the base model results. That leaves two live questions at once: whether users can form problematic bonds with systems, and how well existing evaluations capture that risk. Those questions also sit near this site’s coverage of mourning after AI companion changes and the distinction between AI welfare and AI rights.
OpenAI says it plans to keep working on the category rather than drop it. The under-18 section closes with this statement: “We remain committed to improving how our models support and protect teens.”
Frequently asked questions
Does OpenAI say it measures emotional reliance directly?
Yes. OpenAI’s 7 October system card says its dedicated under-18 evaluations include emotional reliance, and the same card says its dynamic multi-turn benchmarks also include emotional reliance.
Did GPT-6 score worse than GPT-5.6 on emotional reliance?
On the under-18 evaluations, yes. OpenAI says both October GPT-6 models showed statistically significant regressions on emotional reliance versus their GPT-5.6 counterparts. On the dynamic multi-turn emotional-reliance benchmark, OpenAI says the GPT-6 models performed comparably to GPT-5.6 and that none of the observed differences was statistically significant.
Does the system card say these scores reflect ordinary ChatGPT use?
No. OpenAI says the under-18 and dynamic evaluations were deliberately built around difficult cases and long-tail risks, and it says the results should not be interpreted as estimates of how frequently those behaviours occur in typical production use.
How does OpenAI explain the emotional-reliance regression?
OpenAI says it reviewed the regression and found that the evaluation is overly sensitive to benign nicknames such as bro or bestie when users explicitly request them. The company also says some teen protections are implemented through an additional classifier-based block that is not captured in the reported under-18 results.
Caveats: The score changes, significance judgments, and explanations discussed here come from OpenAI’s own 7 October system card. OpenAI also says its under-18 and dynamic evaluations use difficult cases and do not estimate average ChatGPT traffic.