The timing of this article on AI hallucinations feels almost cosmic! I wrote about this exact topic earlier this week(monday), exploring the untapped potential of these "creative inaccuracies" as I call them.
What resonates most with me in your piece is the shift from seeing hallucinations as fatal flaws to viewing them as collaborative opportunities. I've found this exact mindset shift transformative in my own work. The distinction between automation and augmentation is crucial - when we collaborate WITH these systems rather than expecting them to work FOR us, the relationship fundamentally changes.
I've been experimenting with those "impossible combinations" techniques he mentions but for product innovation rather than coding. What's fascinating is how often the hallucinated connections lead to genuinely novel approaches that wouldn't emerge from purely factual thinking. It's like having a brainstorming partner who isn't constrained by conventional wisdom or industry assumptions.
That said, I completely agree with the accountability point. The moment we abdicate responsibility for verifying outputs is when things get dangerous. I believe the most sophisticated AI users aren't those avoiding hallucinations entirely, but those who know when to embrace creative exploration versus when to demand strict factuality. That "hallucination awareness" makes all the difference between a productive partnership and a risky dependency.
Here is a lesson from the 1981. Cadillac built an engine that could shut down two or four cylinders of a V8 engine to enhance fuel economy. It was a great idea and was called the 4-6-8 cylinder deactivation engine. Problem was that the environment under a hood can be wet, greasy, and hot--not a great environment for electrical sensors and connectors. So you might be pulling out from a stop at a street light only to experience your motor going from 8 cylinders to 4. The engine only lasted one year but was ahead of its time.
The bottom line is that Detroit learned that advanced technology (yes, this applies to AI as well) is best when it is used to optimize or improve an already sub-optimal but reliable system. If the advanced technology fails completely, you still have a reliable but sub-optimal system.
I love my ProWritingAid grammar checker. It finds sentences with passive tense and offers several alternative sentences in active tense. Some alternatives may not express the same I idea I was seeking but I can pick the one that does. I become the reliable but sub-optimal system!
> Hallucinations cannot be tolerated in critical tasks such as medical diagnosis
Couldn't agree more. As much as it's incumbent on a programmer to verify the code AI is writing, it's far more important for a doctor to verify the diagnoses that an AI is making. I might even go so far as to say we shouldn't be building any tools where AI is solely responsible for diagnosis or decision making, merely assisting doctors doing a task.
"Honestly, AI is like working with people far more than we realize."
Yes, this! It's not like humans ourselves aren’t free of hallucinations, and we know that whenever we work with other people. I suppose with humans, we've evolved hundreds of signals, besides the content of our words, to gauge trustworthiness.
I'd add a qualifier in that "Hallucinations are fine, as long as you're aware of them."
The risks of hallucinations that companies try to eliminate are because, invariably, you'll have AI non-users eventually interacting with LLMs inside broader products. And if you expect them to function like any other software and return precise, reliable responses 100% of the time, you'll run into issues and disappointments soon enough.
But if you're aware of hallucinations, you get to make conscious decisions about which tasks and purposes you use AI for and how to double-check any critical info, etc.
One thing I tend to tell people is that LLMs are great if you need to get the general gist of a topic or high-level suggestions for a task, strategy, etc. Because at those levels of abstraction, minor underlying hallucinations don't have much of an impact.
It's when you start to zero in on the nitty-gritty details that you need to start paying more attention to individual facts, data points, etc., and making sure to double-check them.
Basically, LLMs are a tool like any other and as long as you're applying it to the right task with the right expectations, hallucinations are indeed completely fine, actually.
To provide a little bit of a counter narrative: I would not put apps like Cursor and Perplexity in the same bucket.
The beauty of code is that if Cursor hallucinates code, you will find out pretty much straight away. Because the code simply won’t run. However, in search and the infosphere more broadly, hallucinations are more troubling. When I use Google, I vet my own information, but when Perplexity serves me up an answer I outsource this vetting to the AI. We also don’t know how often Perplexity hallucinates because they are hardly transparant about this (they have no incentive to be).
The reality is that these answers cannot be trusted, yet regardless of the truth value everything Perplexity will output, it will state authoritatively — because that is AI’s superpower and achilles heel: overconfidence.
> if Cursor hallucinates code, you will find out pretty much straight away.
Only if the hallucinations are so bad that the code fails to run. If it mixes up the business logic (i.e. "this button should add X, but actually multiplies X") and not the syntax, it's far more devilish to find the problems.
While it's true that we don't know how often Perplexity hallucinates, I'd wager that it's much less than GPT-3.5. I think the comparisons to Google (while great marketing for Perplexity) are indeed leading to overconfidence in the product.
Humans are not out of a job anytime soon, LOL
I know a few people who would pay really good money to have hallucinations.
The timing of this article on AI hallucinations feels almost cosmic! I wrote about this exact topic earlier this week(monday), exploring the untapped potential of these "creative inaccuracies" as I call them.
What resonates most with me in your piece is the shift from seeing hallucinations as fatal flaws to viewing them as collaborative opportunities. I've found this exact mindset shift transformative in my own work. The distinction between automation and augmentation is crucial - when we collaborate WITH these systems rather than expecting them to work FOR us, the relationship fundamentally changes.
I've been experimenting with those "impossible combinations" techniques he mentions but for product innovation rather than coding. What's fascinating is how often the hallucinated connections lead to genuinely novel approaches that wouldn't emerge from purely factual thinking. It's like having a brainstorming partner who isn't constrained by conventional wisdom or industry assumptions.
That said, I completely agree with the accountability point. The moment we abdicate responsibility for verifying outputs is when things get dangerous. I believe the most sophisticated AI users aren't those avoiding hallucinations entirely, but those who know when to embrace creative exploration versus when to demand strict factuality. That "hallucination awareness" makes all the difference between a productive partnership and a risky dependency.
If you're interested in my full framework for leveraging these creative inaccuracies for innovation, I wrote about it here: https://thoughts.jock.pl/p/ai-hallucinations-creative-innovation-framework-2025
Here is a lesson from the 1981. Cadillac built an engine that could shut down two or four cylinders of a V8 engine to enhance fuel economy. It was a great idea and was called the 4-6-8 cylinder deactivation engine. Problem was that the environment under a hood can be wet, greasy, and hot--not a great environment for electrical sensors and connectors. So you might be pulling out from a stop at a street light only to experience your motor going from 8 cylinders to 4. The engine only lasted one year but was ahead of its time.
The bottom line is that Detroit learned that advanced technology (yes, this applies to AI as well) is best when it is used to optimize or improve an already sub-optimal but reliable system. If the advanced technology fails completely, you still have a reliable but sub-optimal system.
I love my ProWritingAid grammar checker. It finds sentences with passive tense and offers several alternative sentences in active tense. Some alternatives may not express the same I idea I was seeking but I can pick the one that does. I become the reliable but sub-optimal system!
Hallucinations cannot be tolerated in critical tasks such as medical diagnosis. See https://tomrearick.substack.com/p/funny-flubs.
> Hallucinations cannot be tolerated in critical tasks such as medical diagnosis
Couldn't agree more. As much as it's incumbent on a programmer to verify the code AI is writing, it's far more important for a doctor to verify the diagnoses that an AI is making. I might even go so far as to say we shouldn't be building any tools where AI is solely responsible for diagnosis or decision making, merely assisting doctors doing a task.
Hallucinations are creativity, bullshitting.
Love the way you broke this down Charlie.
Given what the technology is and how it’s used, your point that we’ll never get to 0% is spot on. Chasing it is like chasing the wind.
It’s about understanding hallucinations and how to work with them, not try and correct them.
Honestly, AI is like working with people far more than we realize.
As much as we want to shove it in a nice, neat box and expect it to give a 100% repeatable result, that may be the worst thing it could ever do.
"Honestly, AI is like working with people far more than we realize."
Yes, this! It's not like humans ourselves aren’t free of hallucinations, and we know that whenever we work with other people. I suppose with humans, we've evolved hundreds of signals, besides the content of our words, to gauge trustworthiness.
Yup, this resonates with me.
I'd add a qualifier in that "Hallucinations are fine, as long as you're aware of them."
The risks of hallucinations that companies try to eliminate are because, invariably, you'll have AI non-users eventually interacting with LLMs inside broader products. And if you expect them to function like any other software and return precise, reliable responses 100% of the time, you'll run into issues and disappointments soon enough.
But if you're aware of hallucinations, you get to make conscious decisions about which tasks and purposes you use AI for and how to double-check any critical info, etc.
One thing I tend to tell people is that LLMs are great if you need to get the general gist of a topic or high-level suggestions for a task, strategy, etc. Because at those levels of abstraction, minor underlying hallucinations don't have much of an impact.
It's when you start to zero in on the nitty-gritty details that you need to start paying more attention to individual facts, data points, etc., and making sure to double-check them.
Basically, LLMs are a tool like any other and as long as you're applying it to the right task with the right expectations, hallucinations are indeed completely fine, actually.
To provide a little bit of a counter narrative: I would not put apps like Cursor and Perplexity in the same bucket.
The beauty of code is that if Cursor hallucinates code, you will find out pretty much straight away. Because the code simply won’t run. However, in search and the infosphere more broadly, hallucinations are more troubling. When I use Google, I vet my own information, but when Perplexity serves me up an answer I outsource this vetting to the AI. We also don’t know how often Perplexity hallucinates because they are hardly transparant about this (they have no incentive to be).
The reality is that these answers cannot be trusted, yet regardless of the truth value everything Perplexity will output, it will state authoritatively — because that is AI’s superpower and achilles heel: overconfidence.
> if Cursor hallucinates code, you will find out pretty much straight away.
Only if the hallucinations are so bad that the code fails to run. If it mixes up the business logic (i.e. "this button should add X, but actually multiplies X") and not the syntax, it's far more devilish to find the problems.
While it's true that we don't know how often Perplexity hallucinates, I'd wager that it's much less than GPT-3.5. I think the comparisons to Google (while great marketing for Perplexity) are indeed leading to overconfidence in the product.