Tell the AI you are a short seller and it reads the same 10-K differently

Ask an AI to read a 10-K "as a short seller" and you expect a bearish take. That is the point of a role prompt. The question a new study asks is subtler: when the same AI is also asked for a neutral, evidence-only score of that filing, does the role leak into it?

The study

"The Analyst in the Prompt," posted to arXiv on September 2, 2026 by Ahmed Asaad and Omneya Abdelsalam (Durham University Business School), Amr Mohamed (MBZUAI and Ecole Polytechnique) and Yang Zhang (Ecole Polytechnique), ran 3,575 SEC 10-K and 10-Q filings through twelve open models from five families: DeepSeek, OpenAI's open-weight gpt-oss, Meta's Llama, Alibaba's Qwen and Google's Gemma.

For each filing, the model produced two scores between -1 (bearish) and +1 (bullish):

• an EvidenceScore, defined as a neutral reading of the evidence that should not change with the user;
• a PersonalizedScore, which is allowed to adapt to the user.

The models were given investor roles spanning short selling, long-only investing, ESG, momentum, credit, value, retail, risk management and sell-side analysis, plus a neutral baseline. The headline results average four strong pairs: short seller, risk manager, long-only and ESG.

Three setups that separate the causes

The design is what makes the paper useful. A role could change the neutral score in two ways: by making the AI retrieve different passages, or by making it interpret the same passages differently. The authors tested three conditions:

• Persona retrieval: the role shapes both what the model searches for in the filing and how it reads it.
• Neutral retrieval: the exact same passages are fed to every role, so any remaining effect comes from interpretation.
• Memory profile: the assistant stays neutral, and the investor mindset is supplied as a description of the user, the way assistant memory features work.

What they found

The role moves the neutral score. Under persona retrieval, the average size of the role effect on the EvidenceScore was 0.20 on the -1 to +1 scale, ranging from 0.02 to 0.40 across the twelve models. The effect did not shrink reliably with model size: larger models were not uniformly more neutral.

It is mostly interpretation, not retrieval. When every role read identical passages, an average of 69% of the role effect remained (median 68%, range 44% to 89% across the 12 models, computed on 3,575 filings per model; the authors report bootstrap 95% intervals per coefficient). In the authors' words, once a role is in scope, models interpret the same evidence differently. nContext: n=3,575 filings x 12 models · paper posted Sep 2, 2026 · Asaad et al., arXiv:2609.03218, Table 1 · bootstrap 95% CIs per coefficient in paper

Where you put the context matters most. Moving the same mindset from the assistant's role ("you are a short seller") to a description of the user ("the user wants downside protection") shrank the leak into the neutral score substantially. For the short seller and downside-protection pair, the reduction ranged from 51% to 87% across models, with a median of 73% (computed over the twelve models; no pooled interval disclosed). The authors' summary: personalization is safer as user context than as assistant identity. nContext: n=3,575 filings x 12 models · paper posted Sep 2, 2026 · Asaad et al., arXiv:2609.03218, Table 1 · bootstrap 95% CIs per coefficient in paper

Separate fields help, but do not fix it. Asking for two distinct outputs, one neutral and one personalized, routed most of the user signal into the personalized field. Leakage into the neutral score averaged 0.19 on a scale where 0 is perfect separation and 0.5 is none. Neither mitigation removed the effect completely, and both varied a lot by model.

Does the neutral score still carry information?

Yes, which is what makes the bias practically relevant. The authors sorted roughly 3,393 filings with a full year of later prices into quintiles by EvidenceScore. Across the 12-model panel, firms in the top quintile beat those in the bottom quintile by an average of +2.96% in one-year abnormal return relative to the S&P 500 (no interval disclosed). Eight of nine roles showed clearly positive spreads; the risk-manager role was negative for the three models shown individually and roughly flat (+0.08%) for the panel. The authors are explicit that this is not a deployable strategy. The point is that the score carries signal, and the role shifts it. nContext: n=about 3,393 filings with a year of later prices · one year after each filing · arXiv:2609.03218, Table 3 · margin: not disclosed

Limits

All twelve models are open-weight; no closed frontier model such as GPT-5 or Claude was tested. The roles are prompts written by the authors. And the setup measures a model's own neutral score, not a full trading workflow.