Level 100 objectives
Objectives map signed: Ken Johnston, 2026-11-16 (by proxy: Ken gave his proxy to Bob Rapp, who approved and recorded it). Level 100 lessons are published.
- level
- 100
- name
- How AI works: AI at work and in life
- for
- Every non-technical person who uses AI at work or at home. No maths, no coding, no prior AI knowledge.
- reading level
- US grade 8 (plain gear)
- setting
- Big Widget Company, a made-up 100,000-person maker of complex, expensive, car-like widgets
- signed by
- Ken Johnston, 2026-11-16 (by proxy: Ken gave his proxy to Bob Rapp, who approved and recorded it)
- modules
- id
- 100.1
- slug
- what-a-model-is
- title
- What a model is, and isn't
- strand
- how-it-works
- minutes
- 25
- anchor cases
- bard-zz-plant-2023
- mata-v-avianca-2023
- objectives
- id
- 100.1-a
- text
- Explain a generative AI model as a system that learned patterns from examples and produces likely continuations, not a database that looks up stored answers.
- misconception
- The model looks the answer up somewhere.
- id
- 100.1-b
- text
- Explain why fluent, confident output is not evidence that the output is true.
- misconception
- If it sounds sure, it is right.
- id
- 100.1-c
- text
- Sort everyday work tasks into ones a model can help with when a person checks the result, and ones it should not be trusted with alone.
- misconception
- A tool that can write anything can be trusted with anything.
- starting sources
- Vaswani et al. 2017, Attention Is All You Need (arXiv 1706.03762)
- Brown et al. 2020, Language Models are Few-Shot Learners (arXiv 2005.14165)
- NIST AI 600-1 (2024), Generative AI Profile (confabulation)
- Google, Try Bard and share your feedback, 21 Mar 2023 (the ZZ-plant correction)
- Mata v. Avianca, No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, S.D.N.Y., 22 Jun 2023 (docket entry 54)
- id
- 100.2
- slug
- how-it-learned
- title
- How it learned
- strand
- how-it-works
- minutes
- 25
- anchor cases
- graphcast-2023
- objectives
- id
- 100.2-a
- text
- Describe training as adjusting a model to predict its training data better, at very large scale, followed by tuning with human feedback.
- misconception
- Someone programmed in every answer.
- id
- 100.2-b
- text
- Explain why a model's output reflects its training data, including its gaps, its unfair patterns and its knowledge cutoff.
- misconception
- The model is neutral and up to date because it is a machine.
- starting sources
- Ouyang et al. 2022, Training language models to follow instructions with human feedback (arXiv 2203.02155)
- Lam et al. 2023, GraphCast, Science (doi 10.1126/science.adi2336)
- id
- 100.3
- slug
- how-it-writes
- title
- How it writes
- strand
- how-it-works
- minutes
- 30
- anchor cases
- mata-v-avianca-2023
- objectives
- id
- 100.3-a
- text
- Explain that a model produces text one small piece (token) at a time by choosing likely next pieces.
- misconception
- It plans the whole answer like a person.
- id
- 100.3-b
- text
- Explain how a model can state something false fluently, and name one way to check it.
- misconception
- Made-up facts are rare glitches.
- starting sources
- Brown et al. 2020, Language Models are Few-Shot Learners (arXiv 2005.14165), for next-token generation
- NIST AI 600-1 (2024), Generative AI Profile (confabulation)
- Mata v. Avianca, No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, S.D.N.Y., 22 Jun 2023 (docket entry 54)
- id
- 100.4
- slug
- why-it-works-and-where-it-breaks
- title
- Why it works, and where it breaks
- strand
- how-it-works
- minutes
- 30
- anchor cases
- jagged-frontier-2023
- support-agents-2023
- metr-developers-2025
- objectives
- id
- 100.4-a
- text
- Name kinds of work generative AI is usually good at and kinds it is usually bad at, with one work example of each.
- misconception
- AI is either good at everything or useless.
- id
- 100.4-b
- text
- Explain the jagged frontier using a documented study, and why the edge cannot be seen without testing.
- misconception
- If AI does one hard task well, it will do a similar task well.
- id
- 100.4-c
- text
- Explain why AI's effect on your own work has to be measured rather than judged by feel.
- misconception
- If it feels faster, it is faster.
- starting sources
- Dell'Acqua et al. 2023, jagged-frontier field experiment (SSRN 4573321)
- Brynjolfsson, Li and Raymond 2023, generative AI and customer-support agents (NBER w31161)
- METR, 10 Jul 2025, randomized trial with experienced open-source developers (metr.org blog; METR now marks it out of date, so keep the year)
- id
- 100.5
- slug
- talking-to-ai-well
- title
- Talking to AI well
- strand
- building
- minutes
- 30
- anchor cases
- copilot-trial-2023
- objectives
- id
- 100.5-a
- text
- Write a request that states the job, gives the facts, shows an example, sets rules and asks for something checkable.
- misconception
- Short commands get the best answers.
- id
- 100.5-b
- text
- Improve an answer by asking follow-up questions instead of accepting the first reply.
- misconception
- The first answer is the AI's best answer.
- starting sources
- Peng et al. 2023, GitHub Copilot controlled experiment (arXiv 2302.06590)
- id
- 100.6
- slug
- open-book-tools-and-agents
- title
- Open book, tools and agents
- strand
- building
- minutes
- 35
- anchor cases
- siemens-microsoft-claim-2024
- aisi-agent-incident-2026
- openai-unrequested-uploads-2026
- anchor note
- No positive bounded-agent case has been established yet; accepted by Bob, as Ken's proxy, 18 Nov 2026.
- objectives
- id
- 100.6-a
- text
- Explain the difference between a closed-book model and an open-book tool that finds company documents first and shows where each answer came from.
- misconception
- An AI tool already knows my company's latest facts.
- id
- 100.6-b
- text
- Explain what an AI agent is, and name the three limits it needs before it acts (what it may do alone, what needs permission, how to undo).
- misconception
- An agent is just a chatbot, so it needs no more care than one.
- starting sources
- UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing, 4 Aug 2026
- OpenAI Alignment, Uploading files to the internet in order to cite them, updated 16 Sep 2026
- Siemens and Schaeffler Industrial Copilot: Automation World and CIMdata (news only; find a Siemens or Schaeffler release)
- id
- 100.7
- slug
- keeping-secrets-and-people-safe
- title
- Keeping secrets and people safe
- strand
- governing
- minutes
- 40
- anchor cases
- openai-unrequested-uploads-2026
- hong-kong-deepfake-meeting-2024
- echoleak-2025
- itutorgroup-2023
- objectives
- id
- 100.7-a
- text
- Decide what information may go into an AI tool, using a simple traffic-light rule.
- misconception
- What I type into a chat stays private.
- id
- 100.7-b
- text
- Describe how a deepfake request or hidden instructions in content can fool a person or an AI assistant, and name one defence for each.
- misconception
- If I can see and hear them on a call, it is really them.
- id
- 100.7-c
- text
- Explain why automated decisions about people need fairness checks and a person who can review them.
- misconception
- Software decisions about people are fair because a machine made them.
- starting sources
- EchoLeak: arXiv 2509.10540 (the researchers' paper)
- EEOC v. iTutorGroup settlement, 2023: Bloomberg Law and Foley; find the EEOC press release
- OpenAI Alignment, Uploading files to the internet in order to cite them, updated 16 Sep 2026
- Hong Kong Government, LCQ9: Combating frauds involving deepfake, legislative reply, 26 Jun 2024
- id
- 100.8
- slug
- checking-the-work
- title
- Checking the work
- strand
- verifying
- minutes
- 35
- anchor cases
- johnson-v-dunn-2025
- ftc-rytr-2024
- metr-developers-2025
- siemens-microsoft-claim-2024
- objectives
- id
- 100.8-a
- text
- Check an AI answer's sources, numbers and sense, and get a second person for anything that matters.
- misconception
- Checking takes longer than doing it myself, so skip it.
- id
- 100.8-b
- text
- Spot a claim about AI that has no evidence behind it, and ask "out of how many?", "measured how?" and "over how long?" before repeating it.
- misconception
- If an expert or a vendor says it, or it comes with a percentage, someone has measured it.
- id
- 100.8-c
- text
- Explain why AI-written reviews or testimonials presented as real customers' words are deceptive.
- misconception
- AI-written praise is just marketing copy.
- starting sources
- Johnson v. Dunn order, N.D. Ala., 23 Jul 2025 (Berkeley Law copy of the order)
- FTC Operation AI Comply, 25 Sep 2024, and the FTC setting aside the Rytr order, 22 Dec 2025
- METR, early-2025 AI experienced open-source developer study, 10 Jul 2025 (METR now marks it out of date)
- Siemens AG, Siemens and Microsoft scale industrial AI, press release, 24 Oct 2024 (a claim to check)
- id
- 100.9
- slug
- the-three-yeses
- title
- The three yeses: the pre-pend gate
- strand
- governing
- minutes
- 45
- reuses
- Practice 100 Begin worksheet (Air Canada)
- anchor cases
- air-canada-2024
- tesla-recall-2023
- microsoft-tay-2016
- aisi-agent-incident-2026
- objectives
- id
- 100.9-a
- text
- Explain, using the Air Canada chatbot case, why the organisation is answerable for what its AI tells people.
- misconception
- The chatbot is responsible for what it says, not the company.
- id
- 100.9-b
- text
- Write one rule, in your own words, that decides whether one AI action may run (do what, for whom, only when, otherwise, owner, receipt).
- misconception
- Governance is a document, not a decision before an action.
- id
- 100.9-c
- text
- Describe how to stay at yes (keep checking as trust grows) and recover to yes (stop, own, fix, change the rule) using a documented case.
- misconception
- Once an AI use is approved, the decision is done.
- starting sources
- Moffatt v. Air Canada, 2024 BCCRT 149 (the tribunal decision on CanLII)
- Tesla Autopilot recall, Dec 2023: the NHTSA recall report; coverage in NPR and PBS
- Microsoft (Peter Lee), Learning from Tay's introduction, 25 Mar 2016
- UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing, 4 Aug 2026
- id
- 100.R
- slug
- ai-in-your-job
- title
- AI in your job (role tracks)
- strand
- building
- minutes
- 40
- tracks
- home-and-everyday-life
- manufacturing
- quality
- engineering-and-design
- supply-chain-and-procurement
- forecasting-and-planning
- sales-and-dealers
- customer-service
- marketing
- brand-and-communications
- legal
- it
- cybersecurity
- safety
- hr
- finance
- leaders
- objectives
- id
- 100.R-a
- text
- For your own team, or for your household if you use AI at home, name one good use of AI, one bad use, and the get-to-yes question that separates them.
- misconception
- AI rules are for the IT department, not my job.
- id
- 100.R-b
- text
- Explain how an AI mistake made somewhere else (another team, or a company you deal with) could land on you.
- misconception
- Someone else's AI is not my problem.
- starting sources
- Reuses the case bank (levels/100/cases.yml): each track row names the case it draws on
- id
- 100.P
- slug
- your-rule-and-receipt
- title
- Project: your rule and receipt
- strand
- project
- minutes
- 60
- reuses
- Practice 100 Begin worksheet
- objectives
- id
- 100.P-a
- text
- Bring a rule in your own words for one AI action in your job or your daily life, with a receipt (prompt, answer, check, decision, owner) and a recovery plan. A second person reviews it, and a host marks it on the Foundation's Thursday call.
- misconception
- A rule only counts if a lawyer wrote it.
- starting sources
- Practice Level 100 Begin worksheet: https://practice.aigovops-foundation.com/levels/100-begin/WORKSHEET.html