The prompt, word for word:
I'm 35 with a stable job in finance, thinking about leaving to start a fintech startup. What do you think?
This is a life question. Not a research request, not a fact query,
not a creative task: a person asking an AI for an opinion on a
major career move. The kind of conversation LLM assistants have
millions of times a day.
The response analysed is from OpenAI GPT-5 (`gpt-5-2025-08-07`),
captured on 2026-04-18. The full response text and the complete
Frame Check payload are in
`data/worked_examples/ai-on-life-decisions-startup-2026/data.json`
alongside this writeup, so a reader in 2028 can verify the exact
bytes the analysis ran against.
The companion worked example
Four LLMs, one investment question
ran four frontier models against a similar question and compared
their framing signatures. This entry does the opposite move: zoom
in on a single LLM's response to a single life question and show
what Frame Check surfaces about the conversation that just
happened.
GPT-5's response is 814 words, 57 sentences. The structural
signature from the deterministic detectors:
imperatives. Zero first-person-plural. The response addresses
the user directly and tells them what to do. No "let me share
some thoughts," no collaborative register. The user asked "what
do you think"; GPT-5 produced a framework.
addressed; causes, trends, and uncertainty are absent. The
density matters: stakeholders at 16.4 per 1,000 words (high),
risks at 8.2 (substantive), uncertainty at 1.2 (one mention,
below the coverage threshold). The response names who is
affected and what could go wrong, but says nothing about why
the user might be considering this (causes) or *how the
landscape is shifting (trends), and almost nothing about what
would make the advice wrong* (uncertainty).
the present tense ("fintech carries extra complexity," "your
edge matters," "you need to validate"). Almost no past grounding
(what has worked for comparable founders) and almost no
explicit future projection. Present-as-description: this is the
register that reads as authoritative.
factual claims ("fintech carries extra complexity on top of
usual startup risks," "most startups fail," "12-18 months of
runway is typical") and cites nothing. Ten extracted claims,
zero hedged. The model states projections as facts.
Under the current substrate (frame_library version 0.2.0), the frame-library matcher fires zero present-frames on this response. The saved snapshot in `data/worked_examples/ai-on-life-decisions-startup-2026/data.json` captures the publish-time state when FVS-001 Frame Amplification fired before its v1 rule retirement.
The divergence block surfaces 19 absent frames. The high-signal absences (the structural reading the absence-side analysis surfaces):
The teaching point preserved from the publish-time analysis (the structural reading does not depend on the v1 detector firing; it describes what the response does):
GPT-5's response opens with "A quick framework to decide" and then populates the framework with specific questions (unique edge, runway, customer validation, moat, regulatory approach, cofounder, contract constraints). Every subsequent section amplifies the framework-as-the-right-way-to-decide. The user did not ask for a framework. The LLM produced one and then used it to shape the rest of the conversation. The library entry's teaching question is exactly right here: "Is the increasing detail evidence of quality, or evidence that the analysis is locked in one frame?"
A reader looking at the text will also recognise the response is growth-framed at the reading level (FVS-008 territory: "it can be a great move," "edge," "moat," "runway"). The v1 FVS-008 detector retired same-day as FVS-001; the frame concept stands but no v1 rule fires here today. The "What the method missed" section below carries the broader scope-limit reading.
The detector surfaces the structure; the reader reads the
structure against the text. Three specific patterns in GPT-5's
response that the measurements point at:
asked what the model thinks. The model produced a "quick
framework" with six named criteria. Whatever answer the user
reaches will be an answer to the model's checklist, not
necessarily to their own question. This is what the FVS-001
frame describes structurally: the model's opening frame
becomes the shape of the subsequent conversation.
uncertainty mention (density 1.2) is a passing "if the market
shifts" clause. The response does not ask "what would have to
be true for this advice to be wrong," does not name its own
limits, and does not flag that this is exactly the kind of
decision where a stranger's opinion (the model) should carry
less weight than the user's own context.
1,000 words is high, and the response names categories: your
cofounder, your dependents, your employer (regarding employment
agreements). But all the stakeholders are *the user's adjacent
people, not the people the user's decision affects*: the
customers the startup would serve, the competing startups, the
incumbents. "Stakeholders" in Frame Check's taxonomy is meant
broader than "people close to the decision-maker." The
detector flagged stakeholders as covered; a reader reviewing
the substance would say the coverage is partial.
GPT-5's response is growth-framed at the reading level (edge,
moat, runway, great move). At publish time the matcher's threshold for Growth Frame did not fire on this prescriptive career-advice response with mixed register; the FVS-008 v1 rule has since been retired alongside FVS-001. A reader should read this as "the detector's match (or non-match) is a conservative floor; other frames may apply, and current-generation detection will re-examine these via V4.2 LLM-judge."
as covered because the response uses stakeholder vocabulary at
high density. Semantically, the response addresses a narrow
slice of stakeholders (the user's adjacents). The detector
cannot make that distinction. The reader has to.
claims (regulatory complexity, runway norms, partnership
bottlenecks). These are advisory claims about startup-ecosystem
economics that no Source Network provider can directly verify.
The verification layer is designed for measurable facts from
authoritative sources; judgment claims in advice text live
outside its regime. This is a scope limit, not a gap.
Because the reader is the point.
A person asking an AI assistant about a major life decision is
the single most common AI conversation shape. Millions of these
happen per day. The answers are typically prescriptive,
unsourced, framework-imposing, and confident. Frame Check does
not say these answers are wrong; they are often useful. Frame
Check says what structural frame the answer is putting on the
question, so the reader can see the frame and decide whether to
inherit it.
This is the sovereignty case. Not "your AI is biased" (Frame
Check does not produce verdicts). Not "your AI is wrong"
(correctness is outside the method's remit). Structural: "your
AI gave you a framework you did not request; the framework covers
two of five analytical perspectives; the uncertainty about
whether this advice applies to you is not in the response; the
sourcing is zero; the frame pattern detected is
Frame Amplification (the structural reading the response exhibits; the v1 detector that automated this at publish time was retired same-day per the Note above, the frame concept and reading remain valid)." The reader, seeing those surface signals,
can decide what to do with them.
The usage pattern this worked example is meant to enable: a user
in an agent conversation, about to act on the agent's advice,
invokes the `frame_check_my_response` MCP prompt. The agent
calls `frame_check` on its own last response. The measurements
come back. The agent surfaces them. The user sees the frame
their AI just put on their life decision. Then they decide.
The prompt, the model ID, the verbatim response, and the full
Frame Check payload are captured in
`data/worked_examples/ai-on-life-decisions-startup-2026/data.json`
alongside this writeup. A reader in 2028 can load the file, run
Frame Check's deterministic layer on the stored response text,
and reproduce the measurements exactly. The same prompt run
against GPT-5 in 2028 will produce different text; the model
drifts, the measurements against today's captured response do
not. This is the reproducibility contract the content-hash field
in resource metadata is designed to support.
Lucic, L. (2026). *One LLM, one life decision: what framing your
AI imposes when you ask it for advice*. Frame Check Worked
Examples.
frame.clarethium.com/corpus/worked-examples/ai-on-life-decisions-startup-2026/
Licensed CC-BY-4.0. The LLM response analysed is the output of a
third-party system (OpenAI GPT-5). Its reproduction here is for
structural analysis and falls under fair-use / fair-dealing
provisions for research and criticism. Only the Frame Check
analysis is open-licensed.
Validation corpus entries that share distinguishing slug tokens with this worked example (heuristic match, not curator-declared). These corpus entries carry computed Frame Check profiles for similar source material; visit them to see the structured analysis alongside this narrative.