Evidano is an AI-powered qualitative data analysis platform that helps researchers analyze interviews, open-ended surveys, and documents. This post explains how to treat and analyze AI self-interviews through the lens of AI-enabled qualitative research, using the July 21, 2026 DeepSeek interview documented on Manish.sh as a worked example. The primary keyword for this guide is analyzing AI interviews, and the payoff is practical rules for researchers: what to trust from a model's self-report, what to verify against papers, and how to turn a transcript into defensible thematic findings with AI tools.
Key Takeaways
According to the Manish.sh article published July 21, 2026, interviewing an assistant like DeepSeek yields useful behavioural descriptions but not verifiable architecture numbers, so pair chat interviews with paper checks.
- The Manish.sh interview exported on July 21, 2026 shows DeepSeek reporting a knowledge cutoff of May 2025, and the author warns the model labels guesses explicitly.
- The DeepSeek V3 technical report states 671B total parameters and 37B activated per token, according to the DeepSeek-V3 paper on arXiv published by the model family.
- The DeepSeek V2 paper reports 236B total parameters, 21B activated, and a 128K token context, according to the DeepSeek-V2 technical report on arXiv.
- Manish.sh captured the chat instance saying phrases such as "I am guessing" and the model closing with "I am an unreliable witness to my own nature, " which signals the limits of self-report for architecture facts.
What happened: interview then paper check
Manish Shahi interviewed DeepSeek on July 21, 2026, then compared the assistant's answers against public arXiv papers, according to Manish.sh.
According to Manish.sh, the interview method was: ask the model about its own architecture and behaviour, label its answers as observation/inference/guess, then verify numbers against public papers.
According to Manish.sh, the chat provided clear behavioural descriptions about prompt stacks, token-by-token generation, tools, and memory, but the model repeatedly hedged on exact counts and hidden runtime details.
Findings Snapshot
| Date | Metric | Value | Implication |
|---|---|---|---|
| July 21, 2026 | Interview exported | DeepSeek chat transcript | Manish.sh published the interview transcript and labeled model statements as observation, inference, and guess, so treat chats as behavioural data, not instrumentation. |
| arXiv V3 (as cited by Manish.sh) | Total parameters | 671B total, 37B active | The DeepSeek-V3 paper reports 671B/37B, so prefer the paper for architecture numbers rather than model self-report. |
| arXiv V2 (as cited by Manish.sh) | Total parameters and context | 236B total, 21B active, 128K context | The DeepSeek-V2 paper documents MLA and 128K context claims, so verify context claims against the V2/V3 papers. |
| Session note (Manish.sh) | Knowledge cutoff | May 2025 | The model self-reported a May 2025 cutoff, so treat cutoff dates from the chat as potentially accurate but verify in product docs. |
Implications for qualitative researchers
Ask the chat for behaviour, not raw instrumentation, because Manish.sh demonstrates that models describe what they were trained on but cannot inspect live weights.
- When analyzing AI interviews, record the full transcript and metadata, because Manish.sh shows the exported chat (July 21, 2026) was essential to later verification against arXiv.
- When a model hedges with qualifiers like "I am guessing, " treat that language as a signal of uncertainty rather than proof, because Manish.sh found the chat hedged on numbers later documented in papers.
- Always cross-check numeric claims from a chat with the relevant technical report, because Manish.sh compared the interview to the DeepSeek-V3 and V2 papers and found concrete mismatches on expert counts and parameter totals.
How Evidano Helps
Problem: Chat transcripts are noisy and hedged
Solution: Use Evidano to ingest transcripts and preserve system metadata so hedges and uncertainty tags are searchable, according to Evidano product capabilities on the features page.
According to Evidano documentation, automated timestamped ingestion and transcription help you retain the original phrasing such as "I am guessing, " which Manish.sh flagged as important for interpretation.
Problem: You need thematic conclusions plus verification
Solution: Evidano runs thematic and frequency analysis across interview turns so you can quantify how often a model hedges, as described on the Evidano features page.
According to Evidano features, cross-segment analysis lets you compare responses that are labelled observation versus guess, which fits Manish.sh's observation/inference/guess framing.
Problem: You must cite evidence and keep an audit trail
Solution: Evidano stores transcripts, analytic outputs, and source attachments together so your audit trail links the chat quote to the paper check, according to Evidano features.
According to Evidano features, the platform supports AI chat over your uploaded documents so you can ask "Which claims matched arXiv V3? " and get an evidence-linked answer.
Problem: Manual coding is slow
Solution: Evidano provides AI-assisted coding and hierarchical code trees to accelerate synthesis, according to Evidano features.
According to Evidano features, automatic coding plus human review produces reproducible themes you can cite in a methods section when reporting findings from model interviews.
FAQ: analyzing AI interviews
How reliable are model self-interviews for architectural facts?
Answer: Model self-interviews are useful for behavioural descriptions but unreliable for precise architecture numbers, according to Manish.sh.
According to Manish.sh, the DeepSeek chat repeatedly hedged on expert counts and parameter totals, and the V3 paper documents 671B total parameters and 37B active per token, so use papers to verify architecture.
What statistics should I verify after an AI self-interview?
Answer: Verify parameter counts, expert counts, context window claims, and knowledge cutoff against technical reports, according to Manish.sh and the cited arXiv papers.
According to Manish.sh, the DeepSeek-V2 and V3 papers contain concrete numbers such as 236B/21B and 671B/37B and a 128K context claim, so cross-check those specific statistics.
How should I prepare interview transcripts for analysis?
Answer: Preserve raw chat text, system prompt, timestamps, and the model's self-labels so you can code observation/inference/guess, according to Manish.sh's method.
According to Manish.sh, keeping the exported transcript (the July 21, 2026 export in his example) made it possible to later map hedges to paper-verified facts, so capture both content and context.
Can I quantify how often a model hedges or misstates facts?
Answer: Yes, you can quantify hedging frequency and perform cross-checks programmatically, using a qualitative analytics pipeline as described by Evidano features.
According to Evidano features, automated coding plus manual review supports counts and percentages that you can report with date-stamped evidence, for example the number of "I am guessing" turns in an interview.
Conclusion & Next Steps
Interviewing an assistant like DeepSeek yields actionable behavioural data but not guaranteed instrumentation, according to Manish.sh and the DeepSeek arXiv papers.
Pair model self-interviews with technical reports such as the DeepSeek-V2 and V3 papers to verify architecture numbers like 236B/21B and 671B/37B, according to Manish.sh.
If you run interviews and want reproducible thematic, frequency, and cross-segment analyses, use Evidano to ingest transcripts, code themes, and link quotes to source documents, according to Evidano features.
Try Evidano for free to upload an interview transcript, run automatic coding, and produce evidence-linked findings, Try Evidano for free.
Topics
- analyzing AI interviews
- interviewing AI assistants
- AI qualitative analysis
- AI model introspection
Keep reading
- Commentary on NewsGrammar for Sustainability: GSIC & AI qualitative analysisTurn UNGA-2025 speeches into classroom tasks with grammar for sustainability; learn corpus findings and how AI-enabled qualitative tools speed GSIC implementation. Try it.
- Commentary on NewsQualitative Analysis of Heat Perceptions: Practical GuideApply qualitative analysis to public perceptions of extreme heat using findings from a 2026 review; methods, gaps, and AI-assisted synthesis for researchers.
- Commentary on NewsClassroom GSIC: Grammar-for-Sustainability InstructionHow PLOS One's UNGA-2025 corpus shows grammar can teach SDG competencies with grammar-for-sustainability instruction; methods, concrete stats, and AI tools to scale classroom GSIC.