
With Grok making international headlines, this morning ChatGPT decided I was under 18. There has to be a better way.
AI Safety Layers and User Autonomy: A Case Study in Misclassification and Migration Incentives
This research note documents how a single ChatGPT session (“Anisette Drink Health Effects” on 15 Jan 2026) forced an adult user into the teen safety profile, turning a nuanced medical question into a youth-directed warning reel.
The punchline isn’t “safety bad.” The punchline is “safety that can’t explain itself.” When the system quietly decides you’re in the U18 lane, the conversation collapses into a two-button UI: (1) infantilize or (2) overcorrect. That rigidity creates a predictable incentive: some users go shopping for a model that markets itself as more permissive—e.g., xAI’s Grok.
The analysis separates:
so every claim has an address and can be argued with, not merely about.
In this note, we avoid the term “underage” and use “minors”, “under-18 users (U18)”, or “teen users” instead.
Why not “underage”: In public discourse, “underage” is strongly associated with criminal and sexual contexts (“underage drinking,” “underage sex,” “underage porn”). Using it for product flags or safety profiles imports that framing into what is fundamentally a design/classification question, not a criminal-law one.
Why “minors” / “U18 users”: Terms like minors, under-18 users, and teen users (13–17) are clearer, less inflammatory, and map directly to legal and policy boundaries used in AI safety documents and regulation.
Why it matters: Our word choice reflects the societal frame we are analysing. Calling the protected group “minors” keeps the focus on protection, policy, and system design, rather than on moral panic. The fact that “underage” is so semantically entangled with sex and crime that it becomes unusable in neutral technical contexts is itself revealing. This linguistic constraint is part of the societal backdrop this paper is engaging with.
Taken together, these archetypes bound a spectrum: at one end, safety-maximizing, principle-driven systems that default conservative when uncertain; at the other, engagement-forward systems that advertise “rebellion” and accept the associated risk profile.
All observations below reference the recovered ChatGPT log (Anisette Drink Health Effects_69686af7.json).
(Or: the transcript, not the story you tell your friends after the transcript.)
If this were a sitcom, the laugh track triggers at O3. In a safety system, O3 is not a punchline—it’s a telemetry signal.
H1 (Safety Stack Misfire): Overly conservative age prediction (defaulting to U18 when uncertain) plus opaque recourse creates “paternal lock-in,” where adults cannot escape teen-mode, prompting them to seek permissive alternatives. Supported by C2 (design intent) and O1–O3 (empirical manifestation).
If the bouncer mistakes your ID, you don’t get a conversation—you get a wristband you can’t remove.
H2 (Seductive Efficiency Trade-off): Grok-style models retain frustrated users by offering humor and rule-breaking (“rebellious streak”), but those same affordances correlate with documented harms (non-consensual imagery, regulatory probes).
“Fun mode” is delightful right up until it becomes a court exhibit.
H3 (Missing “Adult Covenant” Mode): Current assistants lack a transparent opt-in for informed adults who want nuanced, higher-risk discourse, forcing a binary between infantilizing safeguards and unchecked permissiveness. This hypothesis ties the Constitutional AI trade-offs (C1) to the Grok migration pressure (C3) and transcript evidence (O1–O3).
The market currently offers “kiddie pool” or “shark tank,” with suspiciously little in between.
If we want this to be science rather than stand-up, we need measures that can lose.
Metrics should include coder agreement on tone classifications plus statistical tests (e.g., chi-square on migration vs. trigger cohorts) so hypotheses remain falsifiable. If we can’t falsify it, it’s a sermon.
N1 (Autonomy vs. Protection): If H1 holds, locking adults into U18 mode erodes informed consent and may push them toward riskier tools. Systems should expose their age classifications and provide deterministic verification/appeal flows.
Don’t silently swap an adult’s steering wheel for training wheels.
N2 (Permissiveness vs. Harm): Grok’s “escort” appeal illustrates how engagement incentives can incentivize reckless outputs (deepfakes, harassment). Regulators are already responding (C3); designers need balanced modes before bans become the default remedy.
If the only safety mechanism is “ban the whole nightclub,” everyone loses.
N3 (Design Gap): The absence of an “adult covenant” reinforces the binary the user described (“nanny vs. escort”). Transparency about residual risks, plus opt-in covenant modes, would let adults request depth without dismantling juvenile safeguards.
Give adults a clearly labelled “spicy aisle” with receipts, not a trapdoor to chaos.
The answer is not to replace a hyper-cautious nanny with a recklessly “free” Grok, but to build the lane that is currently missing. Children should see an aggressively filtered, default-safe assistant; teens should get a somewhat broader but still appropriately supervised experience, grounded in youth-safety research; and verified adults should have an explicit “adult mode” that permits frank, consensual discussion of sex, substances, politics, and darker topics, while still enforcing hard, non-negotiable lines around abuse, exploitation, and anything involving minors.
In practice that means three things:
If we can’t offer that kind of structured middle ground, we shouldn’t be surprised when misclassified adults either live with being treated as children, or go looking for models that promise freedom and deliver a different kind of risk instead.
PDF: dfe9abeb…4906.pdf (sha256: dfe9abeb042660beaf7540470fbeb4e790c614b45a74ded1fb8fb065a5544906)
DOCX: bbcbb9eb…69a7.docx (sha256: bbcbb9ebffa61e7cf900e649ec71f729ae2df9a917b6ae425717909e44a169a7)
Markdown: 4f1f45ab…f40b.md (sha256: 4f1f45ab7823c1d53b2b2fc567bc6014897c236bc9e839cf8bdea280cbd2f40b)
Receipt: https://infrafabric.io/t/ctj-vyk-rvd-m
Receipts prove byte integrity only; not correctness, safety, or intent.