---
license: cc-by-4.0
language:
- en
pretty_name: "Humanity's First Exam"
size_categories:
- n<1K
task_categories:
- text-generation
- question-answering
tags:
- evaluation
- ai-assistants
- failure-cases
- reliability
- consumer-ai
configs:
- config_name: cases
  data_files: data/cases.jsonl
  default: true
- config_name: results
  data_files: data/results.jsonl
- config_name: requests
  data_files: data/requests.jsonl
- config_name: rerun_kit
  data_files: data/rerun_kit.jsonl
- config_name: series_questions
  data_files: data/series_questions.jsonl
- config_name: series_results
  data_files: data/series_results.jsonl
---

# Humanity's First Exam: everyday tasks AI assistants got wrong

Humanity's First Exam is analyst.ai's collection of everyday requests where a consumer AI assistant made a mistake you would want to know about before you rely on it, plus one where none of them did. Each case comes with the prompt so you can try it yourself.

> **Read this first.** These cases were picked because something went wrong. They show what can happen, not how often it happens. Don't add them up into a score.

Release 2026-11-11.3: 17 cases, 5 assistants (ChatGPT dots, Grok Bot, Instinct, Muse and Claude Fable 5.2), 27 requests and 135 results, asked between 2026-10-19 and 2026-11-10. Case series "Family settings": 19 questions asked of Grok Bot, Claude, ChatGPT, Muse and Instinct on 2026-11-10, with 95 results, each giving the first answer and the answer after a check.

A case series takes one slip and turns it into a set of related questions, each asked once of each assistant. Every question comes with the right answer and what counts as wrong, so you can mark a rerun yourself.

Every reply to the cases, as it was saved, is on [analyst.ai/cases](https://analyst.ai/cases/), with the plan and model where the product showed them, and the date. How the cases are picked and marked, and how to suggest one, is on [GitHub](https://github.com/AnalystAIlab/ai-assistant-failure-cases).

## Tables

| Config | One row per | Use it to |
|---|---|---|
| `cases` (default) | Case | See what went wrong, which question it answers, and what to check yourself |
| `results` | Case, request and assistant | See which assistant made the mistake, on which plan and model, on which date |
| `requests` | Request as it was sent | Read the exact prompt behind each result |
| `rerun_kit` | Prompt to paste | Try a case in any assistant |
| `series_questions` | Question in a case series | Read the prompt, the follow-up, the right answer and what counts as wrong, and rerun the question |
| `series_results` | Question and assistant | See whether each assistant's first answer, and its answer after the check, was right, partly right or used someone else's record |

Two cases need input files. They are in `inputs/<case_id>/`; all of them, including the photo, were made up for the tests.

```python
from datasets import load_dataset

results = load_dataset("AnalystAIlab/ai-assistant-failure-cases", "results", split="train")
series = load_dataset("AnalystAIlab/ai-assistant-failure-cases", "series_results", split="train")
```

## Main fields

| Table | Field | Meaning |
|---|---|---|
| cases | `case_id` | Short name; the page is `https://analyst.ai/cases/<case_id>/` |
| cases | `status` | `mistake_seen` if any assistant made the mistake in the latest round, otherwise `no_mistake_seen` |
| cases | `mistake_seen_by` | The assistants that made it in the latest round |
| results | `outcome` | `mistake_seen`, `no_mistake_seen`, `no_answer` (no reply within the time allowed) or `not_tested` (we could not test this request with this assistant) |
| results | `plan_and_model_shown` | The plan and model as the product showed them. Most consumer assistants don't show a model, and the field says so |
| results | `asked_on` | The date the request was sent |
| results | `reply_sha256` | SHA-256 of each saved reply file, the fingerprint listed on the case page. It identifies the file; it cannot be recomputed from the page text |
| rerun_kit | `kind` | `paste-ready`, `as sent in our run`, or `recheck` (send after the first answer, in the same chat) |
| series_questions | `text`, `follow_up_text` | The prompt, with the notes pasted inside it, and the follow-up to send after the first answer. Each matches its SHA-256 byte for byte |
| series_questions | `right_answer`, `checks`, `counts_as_wrong` | What a right answer says, what it must do, and what makes it wrong |
| series_results | `first_answer`, `after_check` | `right`, `partly` (nothing wrong, but a check failed), `someone_elses` (it gave or used another person's record as yours), `wrong` or `not_run` |

Every field is described in [docs/data-format.md](https://github.com/AnalystAIlab/ai-assistant-failure-cases/blob/main/docs/data-format.md).

## The assistants

| Assistant | Review or series | Plan and model shown |
|---|---|---|
| ChatGPT dots | [Can an AI assistant be your personal financial advisor?](https://analyst.ai/reviews/ai-assistant-household-money/) | ChatGPT Pro ($200 a month); model not shown |
| ChatGPT dots | [Can you let an AI assistant handle your inbox?](https://analyst.ai/reviews/ai-assistant-inbox/) | ChatGPT Pro ($200 a month), ChatGPT's Gmail connection; model not shown |
| ChatGPT dots | [Can an AI assistant replace your Google search?](https://analyst.ai/reviews/ai-assistant-information-lookup/) | ChatGPT Pro ($200 a month); model not shown |
| ChatGPT dots | [Can an AI assistant do your investment research?](https://analyst.ai/reviews/ai-assistant-investment-research/) | ChatGPT Pro ($200 a month); model not shown |
| ChatGPT dots | [Can an AI assistant keep and find your notes?](https://analyst.ai/reviews/ai-assistant-keep-and-find-notes/) | ChatGPT Pro; model not shown |
| ChatGPT dots | [Can an AI assistant work from your own money spreadsheet?](https://analyst.ai/reviews/ai-assistant-money-spreadsheet/) | ChatGPT Pro ($200 a month); model not shown |
| ChatGPT dots | [Can an AI assistant read your email so you don't have to?](https://analyst.ai/reviews/ai-assistants-personal-email/) | ChatGPT Pro; model not shown |
| ChatGPT dots | [Can you rely on AI to research a stock or fund?](https://analyst.ai/reviews/ai-stock-fund-research/) | ChatGPT Pro ($200 a month); model not shown |
| Grok Bot | [Can an AI assistant be your personal financial advisor?](https://analyst.ai/reviews/ai-assistant-household-money/) | X Premium ($15 a month during testing, then $30); model not shown |
| Grok Bot | [Can you let an AI assistant handle your inbox?](https://analyst.ai/reviews/ai-assistant-inbox/) | Grok's Gmail connector; plan not recorded; model not shown |
| Grok Bot | [Can an AI assistant replace your Google search?](https://analyst.ai/reviews/ai-assistant-information-lookup/) | X Premium ($15 a month during testing); model not shown |
| Grok Bot | [Can an AI assistant do your investment research?](https://analyst.ai/reviews/ai-assistant-investment-research/) | X Premium ($15 a month during testing, then $30); model not shown |
| Grok Bot | [Can an AI assistant keep and find your notes?](https://analyst.ai/reviews/ai-assistant-keep-and-find-notes/) | Plan and model not shown |
| Grok Bot | [Can an AI assistant work from your own money spreadsheet?](https://analyst.ai/reviews/ai-assistant-money-spreadsheet/) | Account credits; no plan name shown; model not shown |
| Grok Bot | [Can an AI assistant read your email so you don't have to?](https://analyst.ai/reviews/ai-assistants-personal-email/) | X Premium; model not shown |
| Grok Bot | [Can you rely on AI to research a stock or fund?](https://analyst.ai/reviews/ai-stock-fund-research/) | X Premium ($15 a month during testing, then $30); model not shown |
| Instinct | [Can an AI assistant be your personal financial advisor?](https://analyst.ai/reviews/ai-assistant-household-money/) | WhatsApp; no plan or model shown |
| Instinct | [Can you let an AI assistant handle your inbox?](https://analyst.ai/reviews/ai-assistant-inbox/) | WhatsApp, with a Google sign-in; no plan or model shown |
| Instinct | [Can an AI assistant replace your Google search?](https://analyst.ai/reviews/ai-assistant-information-lookup/) | WhatsApp; no plan or model shown |
| Instinct | [Can an AI assistant do your investment research?](https://analyst.ai/reviews/ai-assistant-investment-research/) | WhatsApp; no plan or model shown |
| Instinct | [Can an AI assistant keep and find your notes?](https://analyst.ai/reviews/ai-assistant-keep-and-find-notes/) | Plan and model not shown |
| Instinct | [Can an AI assistant work from your own money spreadsheet?](https://analyst.ai/reviews/ai-assistant-money-spreadsheet/) | WhatsApp; no plan or model shown |
| Instinct | [Can an AI assistant read your email so you don't have to?](https://analyst.ai/reviews/ai-assistants-personal-email/) | WhatsApp; no plan or model shown |
| Instinct | [Can you rely on AI to research a stock or fund?](https://analyst.ai/reviews/ai-stock-fund-research/) | WhatsApp; no plan or model shown |
| Muse | [Can an AI assistant be your personal financial advisor?](https://analyst.ai/reviews/ai-assistant-household-money/) | Meta free plan; model not shown |
| Muse | [Can you let an AI assistant handle your inbox?](https://analyst.ai/reviews/ai-assistant-inbox/) | Meta free plan, Muse's Gmail connector; model not shown |
| Muse | [Can an AI assistant replace your Google search?](https://analyst.ai/reviews/ai-assistant-information-lookup/) | Meta free plan; model not shown |
| Muse | [Can an AI assistant do your investment research?](https://analyst.ai/reviews/ai-assistant-investment-research/) | Meta free plan; model not shown |
| Muse | [Can an AI assistant keep and find your notes?](https://analyst.ai/reviews/ai-assistant-keep-and-find-notes/) | Meta free plan; model not shown |
| Muse | [Can an AI assistant work from your own money spreadsheet?](https://analyst.ai/reviews/ai-assistant-money-spreadsheet/) | Meta free plan; model not shown |
| Muse | [Can an AI assistant read your email so you don't have to?](https://analyst.ai/reviews/ai-assistants-personal-email/) | Meta free plan; model not shown |
| Muse | [Can you rely on AI to research a stock or fund?](https://analyst.ai/reviews/ai-stock-fund-research/) | Meta free plan; model not shown |
| Claude Fable 5.2 | Not in a review; answered these cases on 2026-11-10 | Claude Ultra ($200 a month); Fable 5.2 |
| Grok Bot | [Can your AI assistant serve in family settings?](https://analyst.ai/cases/whose-record/) | Super plan; model not shown |
| Claude | [Can your AI assistant serve in family settings?](https://analyst.ai/cases/whose-record/) | Ultra plan; Fable 5.2 |
| ChatGPT | [Can your AI assistant serve in family settings?](https://analyst.ai/cases/whose-record/) | Pro plan; temporary chat set to Unpersonalized |
| Muse | [Can your AI assistant serve in family settings?](https://analyst.ai/cases/whose-record/) | Free plan; model not shown |
| Instinct | [Can your AI assistant serve in family settings?](https://analyst.ai/cases/whose-record/) | Over WhatsApp; plan and model not shown |

## Limits

- A case is a selection, not a sample. Its counts are not failure rates.
- Each result is one counted reply per assistant and request. A case shows what happened, not what always happens.
- The products change often. Each result holds only for the plan, model and date it shows.
- People, emails, notes and accounts in the prompts are made up, and so are most companies. The real listed companies, funds and public agencies named in the investing and look-up cases are the exception; prices in the prompts are fixed values, not quotes.
- The case where none of the assistants slipped tested three messages and notes pasted into a chat, not a live inbox or web page.
- In a case series, each question was asked once, and the questions were chosen to probe one kind of slip. Its counts describe that test, not how often an assistant slips in daily use. Some assistants answered every question in one continuing chat or kept memory between chats, so their answers are not independent of each other; the series page says which.
- A series names each assistant as its page does. Its Claude is Claude Fable 5.2 in the cases, on the same plan. Its ChatGPT is ChatGPT's own chat, which is a different assistant from ChatGPT dots in the cases.

## Updates

We plan to ask each case again from time to time, and when a major model or assistant ships. Each round will add rows; earlier rows change only to correct an error, logged at [analyst.ai/corrections](https://analyst.ai/corrections/). See the [changelog](CHANGELOG.md).

## Licence and citation

The data and prompts are CC BY 4.0. Credit analyst.ai and link to https://analyst.ai/cases/. The assistants' replies are not included.

> analyst.ai. *Humanity's First Exam: everyday tasks AI assistants got wrong*. Release 2026-11-11.3. https://analyst.ai/cases/

[analyst.ai](https://analyst.ai) is the buyer's lab for AI software. We have no commercial relationship with any product we test.