Part 1 said models are tracks. Part 2 showed what an AI agent did when it ran without its safeguards. This piece is about the fog around the line.
Doom forecasts sell. Stall stories sell. "Wait for the final model" sells. None of that is an operating system for people putting AI into real systems. Those are incentives wearing a weather report.
Forecasts move money and calendars
Apocalyptic press climate: vendors sell caution, delay, or shields that never touch your ledger. Euphoric climate: autonomy with no gates. "Labs are slowing down" climate: buyers freeze projects that could ship on tracks already in the ground.
Part 1 covered this month's pacing essays. Amodei argued for pacing because progress sped up. Headlines still flattened that into stall talk. MIT Technology Review called the messaging moment a "doomer turn." Fair as mood. Useless as a buy plan. Read the essays yourself.
The flatten is the product. Attention follows fear and finality. Procurement follows attention unless someone demands a verify step.
Fog costs even when nothing has crashed. Scenario meetings stack up. Roadmaps slip. Nobody opens the CRM to see if last month's pilot wrote one clean record.
Three buckets. Only two deserve your week.
Real risk research. Hard questions, methods you can inspect. Evals. Red teams. Failure modes. Model cards. Lab essays. Disagree if you want. Argue with the text.
Operating risk. Closer to the job. Wrong send. Wrong write. Stale retrieval. Tool call to the wrong system. No trail. No human yes before the step that sticks. Ops, IT, and governance live here.
Attention capture. Borrows risk language. Swaps evidence for a round number and a mood. Wants clicks, newsletter opens, or an "AI strategy" budget that never lands a change in a system of record.
The tell is boring and reliable. Primary source for the scary number? Action that changes in your stack this week if it is true? No and no means fog.
One cautionary pair of primary numbers
Job-automation figures show how a number becomes culture.
Frey and Osborne (2013) estimated about 47 percent of US employment sat in occupations their model put at high risk of computerisation across 702 occupations. Their claim: a modelled susceptibility share. Not a verified count of jobs already gone. Not a date for mass unemployment.
Arntz, Gregory, and Zierahn for the OECD (2016) used a task-based method across 21 countries. On average they found 9 percent of jobs automatable on their measure. They argued occupation-level methods can overstate risk because "high-risk" jobs still hold hard tasks.
Same topic. Two primary studies. Different figures. Both get crushed into "AI takes the jobs" or "panic is fake." Neither is a plan for your CRM integration. Neither says whether your agent may send mail without approval.
The rule is simple: no primary source, no percentage. The contrast between these two studies is the point.
Extinction statements are not sprint tickets
On 30 May 2023 the Center for AI Safety published one sentence: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Many lab leaders and researchers signed.
Primary text about priority and scale. Not a backlog item. Not a vote on whether finance needs approve-first this quarter. Mix extinction rhetoric with the vendor shortlist and the fog wins.
Read the serious statements. Keep them in governance. Do not let them replace the dull controls that stop bad writes this week.
How doom and stall turn into a business model
Watch who gets paid.
- Media: absolute language wins the feed.
- Vendors: sell fear (our shield) or destiny (our agent before AGI). Both pitch a forecast, not your verify log.
- Internal politics: forecasts delay a rival or rush a launch without controls.
- Consultancies: roadmaps that never touch systems of record, because the slide is the deliverable.
Bad faith is optional. Incentives are enough. Treat a forecast as fact and you will buy the story that fits it.
Build differently and that model weakens. Primary sources for numbers. Landed-work definition of done. Approval before irreversible actions. Fog has less to sell.
What to do instead
- Split research from runway. Lab essays inform policy. They do not schedule the sprint.
- Primary sources for numbers. No link, no number. Blog paraphrases do not count.
- Done means landed. Mail left the right mailbox. Field on the right CRM record. Ledger line matches the view you named.
- Gate irreversible actions. Human yes before send or write. Polite prompts are not brakes.
- Verify in systems of record. Model "done" is a claim. Source system is the check.
- Judge vendors by what lands, not by how hard they narrate doom, AGI, or pacing.
Same standard as Parts 1 and 2. Tracks keep changing. Finish line should not.
Fog without the costume
No skull art required. Fog is enough. Attention has a cost. Every hour on an unsourced percentage is an hour not spent wiring approval into send.
Real risk exists. Part 2 was a real incident, with primary write-ups, specific failures, and specific fixes. That is what real risk looks like: evidence you can check and a control you can change. Models fail. Agents mis-route. Weights get misused. Labs argue about pacing for reasons that are not imaginary. Take that seriously: read primary text, build controls. Do not take the loudest forecast as a stand-in for either.
FAQ
Are you saying AI risk research is fake? No. Read it. Argue with it. The target is forecasts as operating system or sales script without evidence you can check.
Pause until pacing debates end? No. Build on tracks that exist. Keep approval and verification. Change models when your evals say so.
Vendor waves a big jobs percentage? Ask for the primary paper. Read what it measured. No paper, drop the number.
Isn't "wait for AGI" just prudence? Sometimes caution. Often a way to avoid shipping gated workflows on models that already draft and tool-call well enough for narrow jobs. Prudence looks like gates. Delay looks like fog.
Next in the series, Part 4: the return is in the stations. Application-layer ROI, and how to judge projects by what lands.
Where to see this in practice
For an example of the application layer (approval before anything is sent or written, and verification in the system where the work lives), see Intelli-Assist.
Sources
- Frey & Osborne, The Future of Employment, Oxford Martin School (2013 working paper): https://www.oxfordmartin.ox.ac.uk/publications/the-future-of-employment
- Frey & Osborne PDF: https://oms-www.files.svdcdn.com/production/downloads/academic/future-of-employment.pdf
- Arntz, Gregory & Zierahn, The Risk of Automation for Jobs in OECD Countries, OECD (2016): https://www.oecd.org/en/publications/the-risk-of-automation-for-jobs-in-oecd-countries_5jlz9h56dvq7-en.html
- Center for AI Safety, Statement on AI Risk, 30 May 2023: https://aistatement.com/
- Dario Amodei, We Must Pace the Frontier, Sep 2026: https://darioamodei.com/post/we-must-pace-the-frontier
- MIT Technology Review, 14 Sep 2026 (secondary context on public messaging mood): https://www.technologyreview.com/2026/09/14/1144048/the-ai-industry-has-taken-a-doomer-turn-what-now/