Skip to content
Vinthony

AI risk literacy

How to read confident claims about AI — what's knowable, what isn't, and how to make personal decisions under uncertainty. Written for adults whose work, money, or family decisions are starting to depend on the answer.

Two shapes of bad thinking

The AI conversation has two failure modes. The first is hype: “Everything is about to change, every job is gone, sign up for our course before it's too late.” The second is dismissal: “It's just autocomplete, nothing to see here, we've been through this with the internet.” Both are wrong, and both are emotionally cheaper than the third position — which is to take the technology seriously, acknowledge the genuine uncertainty, and make calm, asymmetric bets accordingly.

Five categories of AI risk

Lumping all “AI risk” into one bucket is the first analytical mistake. Distinguish:

  1. Substitution risk — tasks (and parts of jobs) being done by AI more cheaply, faster, or at higher quality. Already underway in writing, customer support, basic coding, design, paralegal work, translation.
  2. Misuse risk — humans using AI for fraud, disinformation, surveillance, harassment, deepfake blackmail. Most of these aren't new categories of harm; AI dramatically reduces their cost.
  3. Systemic risk — AI integrated into critical systems (finance, infrastructure, defence) in ways that create cascading failure modes or amplify existing fragility.
  4. Concentration risk — a small number of companies owning the substrate on which a large fraction of the economy runs. Geopolitical, regulatory, and commercial implications follow.
  5. Alignment risk — the speculative category. Highly capable systems acting on objectives misspecified by humans. Real researchers take this seriously; the timescales and probabilities are genuinely contested.

Personal decisions usually depend on the first three. The fourth is institutional. The fifth is over the horizon for most practical purposes — though arguably not for engineers working at frontier labs.

How to read AI claims

Use the AI claim evaluation worksheeton any claim that's going to change your behaviour. The seven questions:

  1. Who is making the claim, and what are their incentives?
  2. What would have to be true for the claim?
  3. What is the timescale — “soon” covers a 6-month-to-20-year range; pin it down.
  4. What does the strongest critic say?
  5. What evidence would change my mind?
  6. What's the cost if I'm wrong?
  7. Could a reasonable person read this and ignore it?

Time horizons matter

Claims that are obviously wrong at one timescale are obviously right at another. “AI will displace 30% of knowledge work” is probably wrong this year, probably right over 20 years, and a coin-flip at 7. When you read a confident prediction, ask: over what window? The answer is often vaguer than the prediction implies.

Personal planning works better at the 3-7 year range. Shorter than that, change is bounded by which projects ship in your company this quarter. Longer than that, no one has a serious view.

Cost-asymmetric responses

The useful question isn't “what's the probability AI substantially reshapes my work?” (no one knows). The useful question is: “what's the cost of acting as if it will, versus not?”

Most personal AI bets are cost-asymmetric in your favour:

Notice these are the same moves you'd make as durable personal-strategy advice anyway. AI risk doesn't require exotic responses; it raises the cost of not having a strategy.

Common mistakes

  1. Treating “AI” as one thing rather than a family of capabilities.
  2. Confusing a hype video with deployment reality.
  3. Forecasting from feelings (excitement, anxiety) rather than from concrete capability deltas.
  4. Reading only one tribe (optimists or doomers).
  5. Skipping the timescale question.
  6. Trying to be “AI-proof” instead of AI-resilient.
  7. Letting the AI conversation displace other risks (geopolitical, financial, health) that are more immediate.

Related

FAQ

Will AI take my job?
It will probably reshape your job in the next 3-7 years more than replace it outright — though specific roles (entry-level writing, basic coding, routine knowledge work) are seeing real substitution now. The honest answer is: parts of your job, yes; the whole role, depends on the role and on you.
How seriously should I take AI doom scenarios?
Some risk scenarios (mass disinformation, autonomous cyberattacks, military misuse, accelerated job displacement) are concrete and arguably already happening. Existential or species-extinction scenarios are speculative — we'd label them “speculative” under the evidence policy. Both deserve attention; neither deserves panic. Useful action looks the same whether the worst case is severe or moderate.
Should I learn to code, or is that pointless now?
Coding is becoming more accessible, not less valuable. AI handles the typing; humans still have to specify what to build, debug what goes wrong, and integrate code with the rest of the world. Learning to code today is closer to learning to write than learning to weave — the underlying skill compounds.
Is using AI at work cheating?
Not in 2026. Most companies will reward you for shipping more, faster, with AI help. The risk isn't using AI; it's shipping low-quality work because you trusted the AI's first draft. Treat AI like a fast junior — useful, but you're responsible for the output.
What about deepfakes and disinformation?
Concrete near-term risk, especially around elections, finance scams, and individual-targeted blackmail. Practical responses: verify before believing, raise your bar for “too perfectly aligned with my biases,” and assume any voice or video can be faked when the stakes are high.
Should I use AI on my kids' education?
Cautiously. AI tutoring has real educational upside; it also has real downsides if it substitutes for the slow, frustrating learning where kids develop reasoning skills. Use it like a calculator: as a tool, not a substitute for thinking.