The AI economist's mindset

Every module in this course rests on one sentence, so start here: AI produces the draft; you own the judgement and the accountability. A large language model — ChatGPT, Claude, Gemini — is a fast, fluent draft-generator with no stake in whether it is right. It will produce a market comment, a data-cleaning script, or a policy paragraph in seconds, and it will do so in exactly the same confident tone whether the figures are correct or invented. The economist's job does not disappear in that world; it moves. Less time typing the first version, far more time verifying, deciding, and putting your name on the result.

What AI does genuinely well

• Drafting: a first pass at a weekly NSE and shilling market update, a briefing note, a covering email, the boilerplate 'methodology' paragraph you rewrite every quarter
• Synthesis: pulling one narrative out of many sources — turning a 40-page CBK Monetary Policy Committee statement, three analyst notes and a KNBS release into a single coherent summary
• Code scaffolding: writing the first version of a pandas script to load, rename and reshape a messy KNBS CSV, or an R chunk to plot inflation against the policy rate
• Reformatting and translation: turning a wide table into long format, rewriting jargon for a lay audience, converting dense bullet points into prose and back
• Literature and document triage: reading thirty abstracts or a long PDF and telling you which three deserve your own attention — a starting filter, never the final read

What AI cannot be trusted to own

• Final numbers: the headline figure in a published note, the coefficient in a model, the total in a county budget table — anything a reader will act on
• Causal claims: whether the July rate rise caused the shilling to steady or merely coincided with it. Causation is an argument you must be able to defend, not a sentence you accept because it reads well
• Policy trade-offs: who gains and who loses from a new county levy, and whether that trade is worth making. These are judgements of value and consequence, not text completion
• Accountability: the model cannot be hauled before a committee, lose a client, or have its name on the front page. You can. Whatever you sign, you own

An amplifier of judgement, not a replacement

It is tempting to frame this as AI versus the traditional economist, with a winner. That is the wrong frame. AI is fast, broad and tireless; it has read more than you ever will and forgets nothing you paste in. Traditional analysis is slow, narrow and expensive; but it owns the causal story, knows which KNBS series changed definition three revisions ago, and can be held to account. Put them together and AI becomes an amplifier of judgement: a careful analyst gets dramatically faster, while a careless one now ships confident errors at speed. A 2023 Harvard Business School field experiment on consultants captured this precisely — AI lifted quality on tasks inside its competence but degraded it on tasks just outside, where the model produced fluent, plausible, wrong answers that the less critical users accepted. The tool magnifies whatever judgement you bring. Bring none and it magnifies zero.

The workflow: draft, verify, refine, own

• Draft — let the model produce the first version fast. Give it role, context and constraints, and tell it explicitly to flag anything it is unsure of rather than smoothing over the gap
• Verify — check every number against its source, every citation against the real document, every causal claim against the actual evidence. This is where your hours should go, and it is not optional
• Refine — feed back what was wrong, tighten the framing, cut the hedging and the invented detail. Two or three focused iterations beat one long, vague prompt
• Own — put your name on it knowing you have checked what matters. If you cannot defend a line under questioning, it does not ship

That workflow implies a simple triage for every task on your plate. Fully automate only the reversible and the boilerplate, where a mistake is cheap and obvious — reformatting a table, converting a date column, drafting a routine internal email. AI-assist-then-verify the broad middle — most drafting, summarising and coding — where the model saves you an hour but a silent error would embarrass you. Keep entirely human anything you will sign: final numbers, causal claims and policy recommendations. Notice that verification appears in two of the three columns. That is the point. Verification is not the tax you pay for using AI; it is the job the AI leaves you to do.

The four failure modes to internalise early

• Hallucination — the model states something false with total fluency. Ask for the CBK policy rate on a specific past date and it may simply invent a plausible number
• False confidence — tone never tracks accuracy. The model sounds equally certain whether it is right or wrong, so you cannot read its confidence as a signal the way you might weigh a nervous junior's hedging
• Fabricated data and citations — asked for a source, the model can produce a convincing one that does not exist: 'KNBS Economic Survey 2023, Table 4.7', a court ruling, a working-paper DOI, all invented but perfectly formatted
• Plausible-but-wrong analysis — the most dangerous, because it is not obviously broken: a reasonable-looking regression specification, a sensible-sounding but incorrect elasticity, a mechanism that reads well and is simply not what the data support

Put it together on a task you might do every Monday: a short market update on the NSE and the shilling for your desk. Start by drafting with guardrails. Notice how the prompt below fixes the role, supplies the real numbers yourself, and forbids the model from inventing any it was not given — every unknown must come back tagged for you to fill in, not quietly guessed.

You are an analyst writing a 150-word Monday market update for an East African trading desk.

Context I am giving you (use ONLY these figures):
- NSE 20 Share Index: closed Friday at 1,842, up 0.6% on the week
- USD/KES: 129.3, broadly flat on the week
- CBK policy rate: 10.75% (unchanged at the last MPC)
- 91-day Treasury bill: 8.9% at the latest auction

Rules:
- Do NOT introduce any number, date, or event I have not given you.
- If a claim needs a figure I have not supplied, write [VERIFY: what is needed] instead of guessing.
- Neutral desk tone. State no forecast as fact; label any outlook as opinion.
- End with a one-line 'what to watch this week'.

Draft the update.

The draft comes back in seconds, and now the real work starts. Read it against your four failure modes: is every figure one you actually supplied, is any causal link ('the shilling held because...') asserted beyond the evidence, has the model slipped in a number or an event you never gave it? A fast way to force this is to make the model audit its own output. Ask it to list every factual claim it made, so you can tick each against a source rather than re-reading fluent prose that hides its own assumptions.

Now list every factual or numerical claim in the update you just wrote, as a table with three columns: (1) the claim, (2) the exact figure or fact it depends on, (3) whether that figure came from the context I gave you or from your own memory.

Flag every row where column 3 is 'your own memory'. Do not defend those claims; just mark them so I can verify each against the primary source myself.