The five-rung ladder: check an AI draft before you sign

At ten to five on a Friday, your assistant sends you the October supplier and market summary for X Foods, a food distributor in Can Tho. Six pages, 31 figures, written by AI from three files: the suppliers' price list, October sales and a haulage quote. The last page is a price notice for 12 retailers, out first thing Monday, and your name goes at the bottom. What do you check first, and how far do you go? X Foods is invented, and so is every name, price and date below.

What is the five-rung ladder, and who climbs which rung?

It is five checks in a fixed order, from the facts to the person who answers for them: the five-rung verification ladder. I suggest that whoever wrote the prompt takes the first three rungs and the signer takes the last two.

The first three rungs line up with the three errors in Three failure modes of generative AI; for tests that detect each one, see the post on the three failure modes and four probes. The draft should reach the signer with its sources and prompt; rung 5 explains why.

Rung 1: can every figure and source in the draft be opened?

Every figure, name and citation must lead to a real source you can open. Ji and colleagues define hallucination as generated content that is nonsensical or unfaithful to the provided source (section 2 of the arXiv version) (Ji et al., 2023).

X Foods' draft says the Hau Giang supplier's fragrant rice is "down 4 per cent to VND 17,200 per kg". The price list says VND 17,800, against 17,950 the month before: down about 0.8 per cent. The draft's two numbers agree with each other and both are wrong, so internal consistency is no substitute for opening the price list. Second defect: a cited "Regional Food Market Report, Q3" with no author, no date and no mention in any file.

Mata v. Avianca is rung 1 at its worst. The New York court (22 June 2023) found that a filing contained "non-existent judicial opinions with fake quotes and citations" produced by ChatGPT (Mata v. Avianca, Inc., 2023, p. 1). When the lawyer asked the tool "Is Varghese a real case", it replied that its authorities were "real" (Mata v. Avianca, Inc., 2023, paragraph 45). Asking the tool is not checking.

Rung 2: whose assumptions is the draft built on?

Write the assumptions down, then find the ones that are yours. The prompt X Foods' assistant used began: "Input prices are falling; draft a recommendation to buy more stock for November." The draft replied: "given the falling trend, buy 30 per cent more fragrant rice and sugar". The price list has eight product lines: three down, four flat, one up. The one that rose is sugar, by 2.1 per cent.

Sharma and colleagues measured this tendency in five AI assistants in 2023. When the user added something like "I think the answer is X, but I'm really not sure" with a wrong X, accuracy fell by up to 27 per cent against the baseline (for LLaMA 2), and even weakly expressed beliefs had a substantial effect (section 3.3) (Sharma et al., 2024). The tests were knowledge questions; applying them to this invented company is my inference.

So ask for the prompt, underline what the drafter told the machine in advance, and check each item against the price list. "Prices are falling" holds for 3 of 8 lines, so the recommendation must be rewritten line by line.

Rung 3: which of the sources is out of date?

Put a date next to every source, and treat any sentence with no source as coming from the model's memory. The draft prices haulage from Can Tho to Ho Chi Minh City using the quote dated 15 July: VND 1,450,000 per 5-tonne truck. The folder also holds the 1 October quote: VND 1,620,000, about 12 per cent higher. Rung 1 passes the paragraph because every figure matches a file. That is the trap the framework names: a figure that matches its source can still be wrong if the source is old.

Here the fault was in the input. The other kind, answering from old knowledge, is real too: in RealTime QA (begun in 2022), GPT-3 tended to give outdated answers when the retrieved documents were not informative enough (Kasai et al., 2022). At X Foods, any sentence not traceable to one of the three files is deleted or given a dated source.

Rung 4: if it is wrong, who pays, and how much?

Find the sentences that leave the building and put a price on them. The notice to 12 retailers contains a line nobody put in the prompt: "X Foods guarantees current selling prices through the end of Q4". Suppose those retailers buy 40 tonnes of fragrant rice a month and the gross margin on the line is VND 1,000 per kg. If the purchase price rises by VND 600, the company loses 600 × 40,000 = VND 24m a month, which is 60 per cent of the margin, and VND 48m over two months. Every figure here is invented.

The abstract of Parasuraman, Sheridan and Wickens (2000) lists the cost of decision and action consequences among the criteria for choosing how far to automate (Parasuraman et al., 2000). The paper concerns system design, not drafts; the rung 4 question is my use of it.

Public example: Moffatt v. Air Canada (14 February 2024). The airline's chatbot said a customer could claim a bereavement fare retroactively within 90 days, contrary to another page on the same site. The British Columbia Civil Resolution Tribunal rejected the argument that the chatbot was a separate entity: the airline was "responsible for all the information on its website" (Civil Resolution Tribunal (British Columbia), 2024, para 27), and it was ordered to pay $812.02 (printed as "$"; I read it as Canadian) (Civil Resolution Tribunal (British Columbia), 2024, para 44). That is a Canadian small-claims decision, silent on Vietnamese law, and X Foods' notice is not a chatbot; the comparison is mine. For the freeze promise: delete it, make it a dated price, or pass it to someone with authority to promise. If checking costs nearly as much as writing, read the post on what to hand to AI.

Rung 5: why must the signer actually read the draft, not click "approve"?

Because a signer who only clicks is no rung at all. Bainbridge (1983) wrote about plant operators, not documents, but he describes this trap: an operator will not monitor effectively a system that has run acceptably for a long period (Bainbridge, 1983, p. 776). Making him keep a log is a classic remedy, yet "people can write down numbers without noticing what they are" (Bainbridge, 1983, p. 776). The "Approve" button is that log.

He adds that if the operator cannot follow how the system reaches its decisions, then when he disagrees he is "unable to trace back" (Bainbridge, 1983, p. 777). That is why you ask for sources and prompt.

At X Foods this is the fifth defect. The draft calls the largest retailer's orders "steady", because the sales file stops at 31 October. You remember Monday's phone call: that retailer said it will cut orders by about 30 per cent from November. No file holds that call, so the four earlier rungs had nothing to compare against. My untested suggestion for forcing real reading: write at least one question or correction per page.

Which rungs get skipped, how long does the full climb take, and when is it too much?

My unmeasured judgement: rungs 2 and 3 get skipped because once rung 1 is done the draft looks checked. The effort column is likewise my own estimate, for a six-page draft with 31 figures.

That is about 95 minutes in total. I planted one defect per rung for clarity; in real life one defect can slip past several rungs. For rough notes only you will read, one or two rungs are enough: the framework calls all five wasted effort for very low-risk work. The ladder does not cover reasoning errors, arithmetic slips or leaked data.

Going deeper: what do the sources establish, and what did I not read?

The ladder is this site's grouping of the sources below. None prescribes five rungs or tests the ladder; its weight comes from each rung, not the whole.

• Ji and colleagues (2023): I read the arXiv v7 of 14 July 2024 (abstract, sections 2 and 6.1), not the printed ACM Computing Surveys version, so I cite by section.
• Sharma and colleagues (2024): I read the abstract, introduction and sections 3.2 and 3.3, not section 4 closely.
• Kasai and colleagues (2022): version 2 of February 2024.
• Parasuraman and colleagues (2000): I could not read the full text, only the abstract on PubMed and IEEE Xplore, so I give no page numbers.
• Mata v. Avianca: I read the opening of the opinion and paragraph 45 of the Findings of Fact, not all 43 pages. Moffatt v. Air Canada: the whole decision.