🎮 The Next Input — Issue #167

Sponsored by

⚡ The Briefing — 60 sec

• Meta will record employees’ keystrokes and use it to train its AI models This is insane. Tracking keystrokes now? Seriously? Why am I even surprised... Reuters says Meta is rolling out software on US employee computers to capture mouse movements, clicks, keystrokes, and screen snapshots to train AI agents.

• SpaceX says it has option to acquire Cursor for $60bn Elon wants your codebase now. Reuters reports SpaceX secured an option to acquire Cursor for $60 billion later in 2026, or pay $10 billion for a strategic partnership instead, as it pushes harder into AI developer tooling.

• Millions of dollars in AI camera fines wiped as drivers appeal AI slop will not result in me getting a ticket. Glad they fought back. 9News reports thousands of WA drivers successfully appealed AI camera infringements, wiping about $2.2 million in fines.

🛠️ The Playbook — The Human Override Engine

Mission
Deploy AI into operational workflows without letting surveillance creep, automated penalties, or vendor overreach outrun human judgment.

Difficulty
Intermediate

Build time
3–5 hours

ROI
Fewer bad automated decisions, better trust, and a much cleaner line between useful AI and AI that starts acting like it owns the place.

0) Why This Matters

Three signals, one pattern.

First, Meta’s employee monitoring push shows how quickly AI ambition can slide into workplace surveillance. Reuters reports the company is collecting staff keystrokes and screen activity on work computers to train AI agents, with no opt-out on company-issued laptops.

Second, SpaceX’s move on Cursor shows where power is consolidating. A rocket company with an AI arm wanting one of the best-known coding products on the market is not subtle. Reuters says the arrangement would either give SpaceX an option to buy Cursor for $60 billion or lock in a $10 billion strategic partnership.

Third, the WA AI-camera appeals are a useful reminder that people will push back when automation gets things wrong in the real world. 9News says successful appeals have already wiped millions in fines.

So the move is not:

• trust the automation

• trust the vendor

• assume the system got it right

The move is:

• keep a human override path

• define where AI can recommend versus decide

• build workflows that can be challenged when the machine gets cute

1) Architecture

2) Workflow

1. Identify one workflow where AI is already recommending, flagging, or making decisions.

2. Separate what the model is allowed to suggest from what it is allowed to execute.

3. Add a human override or appeal path for any action with financial, legal, reputational, or employee impact.

4. Record the evidence used by the AI so wrong calls can actually be challenged.

5. Review where the workflow saves time versus where it creates new risk or resentment.

6. Expand only when the system proves it can be both useful and contestable.

3) Example Prompts

Override Design Prompt

You are designing a human override layer for an AI-assisted workflow.

For the workflow below:
- identify what the AI is allowed to recommend
- identify what the AI must never execute on its own
- identify where human override is mandatory
- identify the top 5 failure modes

Workflow:
[insert workflow here]

Appeal Review Prompt

You are reviewing a disputed AI-driven decision.

Check:
- what evidence the AI used
- whether the evidence is sufficient
- where the AI may have overreached
- whether the decision should be upheld, reversed, or sent for review

Return:
1. decision
2. short reason
3. what should change

Surveillance Risk Prompt

You are reviewing an AI workflow for surveillance creep.

Identify:
- what user or employee behaviour is being captured
- whether the capture is necessary
- what privacy or trust risks it creates
- what data collection should be reduced or removed

Vendor Control Prompt

You are assessing whether a workflow is becoming too dependent on one AI vendor or platform.

For the stack below:
- identify lock-in risks
- identify where leverage is shifting away from the team
- identify fallback options
- recommend whether to expand, hedge, or slow down

Stack:
[insert stack here]

4) Guardrails

• Never let AI penalties or employee-impacting decisions run without a challenge path.

• Keep evidence attached to every automated or semi-automated decision.

• Do not collect behavioural data just because the model team wants more training material.

• Separate AI recommendation from AI authority.

• Treat vendor concentration as an operational risk, not just a product choice.

• If users cannot contest a bad AI decision, the workflow is unfinished.

5) Pilot Rollout — 3 hours

1. Pick one AI-assisted workflow with real consequences if it gets things wrong.

2. Map what data it uses, what decision it makes, and who is affected.

3. Add an explicit human override or appeal step.

4. Create a simple evidence log showing what the model saw and why it responded the way it did.

5. Run 10–15 real cases and track where people would have wanted to challenge the result.

6. Keep only the version that reduces friction without removing recourse.

6) Metrics

• Number of AI decisions with a working override path

• Appeal or override rate

• Time to reverse a bad AI decision

• Percentage of automated actions with evidence attached

• User trust score

• Employee or customer complaint rate

• Vendor concentration across critical workflows

Pro Tip: The fastest way to make people hate an AI workflow is not a spectacular failure. It is a dumb call they cannot appeal.

🎯 The Arsenal — Tools & Platforms

• Reviewer queues · simple human override layers for disputes, approvals, and reversals

• Audit logs · essential if you ever want to explain why the AI did what it did

• Airtable / Google Sheets · lightweight way to track overrides, appeals, and evidence trails · Airtable · Google Sheets

• Platform watchlists · useful when giant players start pulling code, data, or workflows deeper into their orbit, as the SpaceX/Cursor move suggests.

• Appeal-first workflow design · because the WA fines story is a good reminder that “AI said so” is not the same as “case closed.”

Copy-paste prompt block:

You are helping me build a Human Override Engine.

For the workflow below:
1. identify what the AI is allowed to recommend
2. identify what the AI must not execute on its own
3. identify where human override is mandatory
4. identify what evidence must be logged
5. identify the top 5 failure modes
6. propose an appeal or review path
7. define the key metrics to track

Workflow:
[insert workflow here]

Return the answer in markdown with sections for:
- Workflow summary
- AI authority boundaries
- Human override points
- Evidence requirements
- Risks
- Appeal path
- Metrics

💡 Free Office Hours

If your AI workflow is getting powerful enough to annoy people, surveil people, or make decisions people might want to fight, I run free office hours to help design the override layer before things get ugly.

Book here: https://calendly.com

Stop Losing Your Money. It's time to upgrade your trading platform.

Your current trading platform is probably letting you down

• Limited assets (no international stocks, no commodities, no pre-IPO companies)

• Limited ability to short

• Limited access to leverage

• Limited trading hours

Liquid is one of the fastest growing trading platforms, allowing users to trade stocks, commodities, FX, and more 24/7/365 from their phone and computer.

Trading on Liquid is as simple as:

1. Pick an asset

2. Pick long or short

3. Pick your position size and leverage

4. Place your trade

The best part is that Liquid markets never close. So no matter what is going on in the world, you are able to keep your portfolio positioned properly.

Start trading 24/7 today

🕹️ Game Over

AI can absolutely move faster. People still need a way to say, “Hang on, that’s wrong.”

— Aaron Automating the boring. Amplifying the brilliant.

Subscribe: link