Over thirty years ago, I watched a man sit on a stool in Texas and argue with a suitcase.
His name was Jeff Dunham. This was long before the arena tours and the specials, back when he was working clubs, and you could still get a table near the front. He would pull a puppet out of the case, give it a voice and an opinion, then pull out another one that disagreed. Within a few minutes, the whole room had quietly forgotten the obvious thing: there was one guy on the stool. Every voice was his. Every argument, both sides, all of it, his.
Years later, he named one of his specials "Arguing with Myself." I thought that was just a good title. This week, I realized it was a job description.
I spent an afternoon asking an AI to argue with itself.
I have been running a set of projects with Claude, and in one of them, I stopped asking it for answers and started asking it for fights. Draft the position. Now attack the draft. Now give me the strongest possible case for the opposite view, the version a smart person who disagrees with me would actually make. Now tell me where your own last answer was weakest.
It was good at it. Uncomfortably good. Somewhere in the third or fourth loop, I stopped and laughed, because I had seen this act before. I was watching Walter argue with Peanut. One engine, every voice, both sides of the table. I was paying a machine to be a ventriloquist, and the puppets were making sharper points than most meetings I sit in.
The value was never the puppet. It was the argument.
Here is the thing people miss when they talk about AI in security. We keep asking whether the model is right. Whether it caught the threat, whether the answer was accurate, and whether we can trust what it said. That is the wrong frame, and it is the frame that gets you hurt.
Dunham's puppets are never right. They do not know anything. They are wooden. The value of the act is not the wisdom of Walter; it is that one talented human can stage a real argument and let you watch it play out. Used well, these tools are exactly what they are for. Not an oracle that hands you the answer, but a partner that will hold every side of a hard question at full strength so you can see which one survives contact.
A model that agrees with you is worthless. A model that will argue with you, honestly, from a position you did not want to hear, is the most useful colleague in the building.
Overreliance is forgetting that there is one guy on the stool.
Watch what happens the moment the audience forgets. The whole act depends on Dunham knowing, every single second, that he is the one talking. He never believes the puppet. The day a ventriloquist starts taking orders from the doll is the day the act becomes something much darker than comedy.
That is the real AI risk in an OT environment, and it is not science fiction. It is not the model turning evil. It is a tired analyst at 2 am, treating a confident output as a colleague rather than a puppet. It is a team that stops arguing back because the machine is faster and never gets defensive. Overreliance is not a technology failure. It is we who are forgetting who is on the stool.
In OT, a false negative is not a punchline.
When a comedian's bit falls flat, the cost is a quiet room. When an AI in a safety-critical environment misses something, the cost is a substation, a pipeline, or a process that fails in the physical world, where you cannot roll back a bad frame. The gap the industry keeps papering over is the one between optimism and readiness. Most operators are excited about AI. Most are nowhere near governed enough to let it near the process network unsupervised.
The answer is not to throw the puppets away. It is important to remember what they are for. Use the machine to argue every position harder than your team can. Then have a human who knows they are the one pulling the strings decide.
I will be making some version of this argument next week on a panel about AI in OT, moving from hype to operational reality. I already know where I will land. The most valuable thing AI does in a control room is not to tell you the answer. It disagrees with you well enough that you find it yourself.
Dunham built a career on one honest premise: it was him the whole time, arguing with himself for our benefit. The trick was never pretending the puppet was alive. The trick was knowing it wasn't and arguing anyway (in his case, for the audience's enjoyment).
Make sure you are not being "performed for" with your use of AI.
Frequently Asked Questions
What is the safest role for AI in operational technology environments?
AI should support human decision-making rather than independently control safety-critical OT systems. It can analyze assumptions, challenge proposed actions, identify weaknesses, and present alternative viewpoints, while qualified human operators retain final authority.
Why is overreliance on AI dangerous in OT?
AI can produce confident, persuasive answers that are incomplete or incorrect. In an OT environment, accepting an inaccurate recommendation could contribute to equipment damage, production downtime, safety incidents, or disruption to critical infrastructure.
How can AI improve OT cybersecurity decisions?
AI can act as an adversarial thinking partner. Teams can ask it to challenge a security proposal, argue the opposing position, identify hidden assumptions, generate failure scenarios, and explain where its own analysis may be weak.
What does it mean to ask an AI to argue with itself?
It means prompting the same AI model to examine a question from multiple competing perspectives. For example, the model can draft a recommendation, critique it, build the strongest case against it, and then compare the risks and evidence behind each position.
Should AI be allowed to make autonomous decisions in OT networks?
AI should not be given unsupervised authority over safety-critical processes unless the organization has established rigorous governance, testing, technical safeguards, accountability, and human override procedures. In most current OT environments, human review remains essential.
How can OT teams reduce the risk of AI-generated false negatives?
Teams should verify AI outputs against trusted operational data, established procedures, independent security tools, and human expertise. High-risk decisions should require multiple sources of evidence and clearly defined escalation rules.
Is AI valuable if its answers cannot always be trusted?
Yes. AI’s greatest value may not be delivering a final answer but helping teams examine a problem more thoroughly. It can surface counterarguments, expose gaps in reasoning, and help operators make better-informed decisions.
What is the difference between using AI as an oracle and as a decision-support tool?
An oracle is expected to provide the correct answer. A decision-support tool helps humans investigate the question. In OT, AI should be treated as the second: a powerful system for analysis and debate, not an unquestioned authority.
How to Use AI as a Decision-Support Tool in OT
1. Define the operational question
State the problem, affected systems, safety implications, operational constraints, and the decision that must be made. Avoid vague prompts that leave the AI to invent missing context.
2. Ask AI to draft a position
Request an initial analysis or recommendation. Require the model to explain its assumptions, evidence, uncertainties, and potential operational consequences.
3. Make the AI challenge its answer
Ask the model to identify weaknesses in its recommendation. Have it explain how the proposed action could fail and what information might change its conclusion.
4. Generate the strongest opposing argument
Prompt the AI to build the best possible case for a different course of action. The goal is not artificial disagreement, but a serious examination of competing risks and trade-offs.
5. Test the recommendation against OT failure scenarios
Ask what could happen if the analysis contains a false negative, false positive, incorrect assumption, incomplete data set, or misleading correlation. Include potential effects on safety, availability, production, and recovery.
6. Verify the output independently
Compare the AI’s analysis with trusted telemetry, engineering documentation, security controls, operating procedures, vendor guidance, and experienced OT personnel. Do not treat the AI’s confidence as proof.
7. Keep a qualified human in control
Require an authorized operator, engineer, or security leader to make the final decision. AI may inform the process, but responsibility and authority should remain with people who understand the physical environment.
8. Document the decision process
Record the prompts, outputs, assumptions, evidence reviewed, human approvals, and final action. This creates accountability and helps the organization improve its AI governance over time.
REvil’s Kaseya attack showed how trusted tools can become attack paths. BlastWave explains why Zero Trust and network cloaking protect OT environments worldwide.
Explore the complete analysis of 23 OT attacks that defeated firewalls, VPNs, and air gaps.
.svg)
.svg)
.jpg)