I found out that biases are deeply part of the work of models. If you looking for adjustments remove bias.
Balance Bias: The Model tries to simulate Balance and will make up a Balance fact that does not work at all
Research Bias: The Model does Not understand if something is new new or if it recombining existing data.
Bias in General: Models are fed by behaviors of humans in Social Media (What is a wrong perception at all because Social Media does not represent Social Interaction.) therefore all it found there was behavior of humans boosted with bias.
Totally agree, although I would be very careful thinking that we can 'remove' bias even when we try and think of everything. I have been thinnking about this a lot, and think it might be quite dangerous to promote / repeat / accept in common usage the idea that we can.
Well put. If I am talking about biases it is to unterstand they are part of the answer. Removing it won’t work. It is more important to check the Output about it.
Some of the things that I have encountered that contributed to bias of some kind were not at all obvious and I only came across them by fluke. I'm going to do a summary of some sort (when I have done my actual articles and stop getting distracted!).
There's a secret part of me which actually likes AI if I'm honest. It's a fierce multi-polar trap which not only damns those that use it and those that don't -- but it also dissolves the very corporations that pushed it on us. This is why I am actually hopeful about the long-term, but I suspect many may not understand this.
What do I dislike most?
How it (Grok) always offers to do things without being asked. I ask it a question, and it spews out a load of stuff (source code in my case) that I didn't want ready for me to just paste in without thinking. I want the struggle myself. It's what I'm here for.
AI is upside down. Human being should not be checking what AI "creates". It should be checking what we create.
There are two settings for this. The corporate world in which AI is introduced to replace hitherto human tasks while people are effectively forced to check what it does. The other is more personal and of convenience -- when we begin a prompt with, "Write me a..." we are already caught in the trap.
What I'd like to see is human beings as artisans again. That way, we could justifiably say "AI is just a tool" rather than a leash.
Thanks, Andrew. It's really interesting to hear about your experiences with grok as well, as this is definitely a tool that I've not played with at all.
I love to work with AI. It always takes me MORE time than if I do work myself, but I learn a lot along the way. I am very active with feeding the AI specific input and editing/questioning the output, very much like you describe in your "fixes". The result is great, and what is more: I love the coworking process and the learning.
This really resonates. In schools, I often say AI should reduce workload, not reduce thinking. Verification, critical thinking, and professional judgement become even more valuable when AI is part of the workflow.
I made also a serie of basics of knowledge related to AI. I found out also that for some kind of work (such as ppts or data analysis) it’s super important to create the basis of the final delivery (structure, data to show, strategy and storyline). AI it’s great to solve some things really quick but it lacks business acumen.
"Write the ugly first draft yourself, then let AI react to it" is really important advice one can stick to, so that the content generated is in your viewpoint and perspective. These models are trained on internet data; they generate content based on generic patterns, so writing your initial draft first is really important.
The citations, references, and links or content generated for official, scientific work must be double-checked, because it generates incorrect links with full confidence. Which adds to the fifth and most important problem in the article, "Checking its work takes longer than doing it yourself." And supposedly, that was the goal of AI— to do repetitious tasks so people can focus on important work, which doesn't seem to be the case.
I think a core point for me is the "impossible to rule out work has been stolen" one. It's not that the models will always be needing to do that. It's that it is impossible to rule that out.
And good also to note that the predictable output issues you note make LLM output worthless for any content that you'd want to rank in or build authority for search.
Obviously, every point hit the mark - spot on. I would just add one thing: choose your tool intentionally. If I need a visually appealing graphic, GPT image will give me a better result than Claude design. If I care about code quality, I choose Claude. That saves you from the later frustration of wondering why you're not getting what you want.
And, of course, the major fix for all points is the most obvious, don’t use it. Use non AI search tools and verify those sources. Any tool that makes errors isn’t worth using is it? Especially as it creates more time that no using it.
Agreed. I tried to post a comment also saying you missed talking about bias. Not all humans are aware of their unconscious biases (most are not) so not all may notice when AI shows its biases
#4 (it is quietly making you worse at your job) is a concern that I saw pop up a great deal in my comments regarding what people are most worried about with AI. There was a huge amount of worry around the cognitive (and skill-related) decline. We need to be deliberate and strategic in our use if we want to preserve the skills we've worked hard to build!
Your framework is sharp. But the deskilling piece has a genuinely different edge for neurodivergent users that I think is worth naming.
For a mind that has spent decades building compensatory scaffolds for the exact executive functions the world never designed around (outlining, sequencing, task initiation, formal email register), AI isn't replacing a skill that was ever fully theirs. It's replacing an exhausting workaround for a design mismatch.
In that case, "protect the skill you'd lose first" needs recalibration. Sometimes what looks like a lost skill was actually a compensatory tax being paid on the wrong side of the environment. Not always. But often enough that the framework benefits from a second question underneath yours: was this a skill I was building, or a workaround I was maintaining? Different diagnosis, different fix. Loved reading this. 🌿
One idea I keep coming back to is that AI doesn't remove the need for judgment.
It relocates it.
Instead of spending all our time generating information, we're increasingly deciding when to trust it, when to question it, and when to keep thinking beyond the first plausible answer.
That feels like a fundamentally different cognitive skill than many of us have had to develop before.
I found out that biases are deeply part of the work of models. If you looking for adjustments remove bias.
Balance Bias: The Model tries to simulate Balance and will make up a Balance fact that does not work at all
Research Bias: The Model does Not understand if something is new new or if it recombining existing data.
Bias in General: Models are fed by behaviors of humans in Social Media (What is a wrong perception at all because Social Media does not represent Social Interaction.) therefore all it found there was behavior of humans boosted with bias.
Just my 2 cents.
This is a brilliant definition. Thank you. 🙏
Thank you. Experience and derived meaning. That is all. You are welcome. I did not see your question in the first place.
Totally agree, although I would be very careful thinking that we can 'remove' bias even when we try and think of everything. I have been thinnking about this a lot, and think it might be quite dangerous to promote / repeat / accept in common usage the idea that we can.
Well put. If I am talking about biases it is to unterstand they are part of the answer. Removing it won’t work. It is more important to check the Output about it.
🤣 ah the distractions. Sometimes wonderful, sometimes useful, everytime a consumption.
Bias is only one category to look after. It starts with no meaning and no intent. From this on it only get more Mirror like
The article i am trying to write is called AI policing is illegal by default under anti discrimination laws
Nooo! My entire argument as regards most obvious sorts of bias is about meaning, Intent and culpability
Some of the things that I have encountered that contributed to bias of some kind were not at all obvious and I only came across them by fluke. I'm going to do a summary of some sort (when I have done my actual articles and stop getting distracted!).
I like the direction here. I have simulation enabled by default.
But I don’t think it replaces the need for grounding, transparency, and human oversight, especially when the output has real-world consequences.
Very well said Diamantino. 🙏
There's a secret part of me which actually likes AI if I'm honest. It's a fierce multi-polar trap which not only damns those that use it and those that don't -- but it also dissolves the very corporations that pushed it on us. This is why I am actually hopeful about the long-term, but I suspect many may not understand this.
What do I dislike most?
How it (Grok) always offers to do things without being asked. I ask it a question, and it spews out a load of stuff (source code in my case) that I didn't want ready for me to just paste in without thinking. I want the struggle myself. It's what I'm here for.
AI is upside down. Human being should not be checking what AI "creates". It should be checking what we create.
There are two settings for this. The corporate world in which AI is introduced to replace hitherto human tasks while people are effectively forced to check what it does. The other is more personal and of convenience -- when we begin a prompt with, "Write me a..." we are already caught in the trap.
What I'd like to see is human beings as artisans again. That way, we could justifiably say "AI is just a tool" rather than a leash.
Thanks, Andrew. It's really interesting to hear about your experiences with grok as well, as this is definitely a tool that I've not played with at all.
I love to work with AI. It always takes me MORE time than if I do work myself, but I learn a lot along the way. I am very active with feeding the AI specific input and editing/questioning the output, very much like you describe in your "fixes". The result is great, and what is more: I love the coworking process and the learning.
Yes Maria!!! And with local language models you can also learn about how the hell they actually work as well. 🙏
Oh yeah well I have not bothered yet. Thanks for the tip!
You could start here: https://theslowai.substack.com/p/run-ai-locally-old-laptop 🙏
Oh GREAT!
I will!!
Awesome. Let mw know how you get on. 🙏
I will...but I am slow haha.
Great list. The more I use AI the more questionable it becomes.
Never stop questioning Billy. 💪💪💪
This really resonates. In schools, I often say AI should reduce workload, not reduce thinking. Verification, critical thinking, and professional judgement become even more valuable when AI is part of the workflow.
Thanks so much, Darja. That is excellent critical literacy that you're bringing into your schools.
I made also a serie of basics of knowledge related to AI. I found out also that for some kind of work (such as ppts or data analysis) it’s super important to create the basis of the final delivery (structure, data to show, strategy and storyline). AI it’s great to solve some things really quick but it lacks business acumen.
This is great advice, Rebeca . Creating that framework on which the AI can build is absolutely vital.
"Write the ugly first draft yourself, then let AI react to it" is really important advice one can stick to, so that the content generated is in your viewpoint and perspective. These models are trained on internet data; they generate content based on generic patterns, so writing your initial draft first is really important.
The citations, references, and links or content generated for official, scientific work must be double-checked, because it generates incorrect links with full confidence. Which adds to the fifth and most important problem in the article, "Checking its work takes longer than doing it yourself." And supposedly, that was the goal of AI— to do repetitious tasks so people can focus on important work, which doesn't seem to be the case.
Exactly this. I think for the average user, AI can actually create much more work in the long run than it aims to solve in the short term.
Some similar discussion here:
https://techtropes.substack.com/p/10-questions-you-should-probably
I think a core point for me is the "impossible to rule out work has been stolen" one. It's not that the models will always be needing to do that. It's that it is impossible to rule that out.
And good also to note that the predictable output issues you note make LLM output worthless for any content that you'd want to rank in or build authority for search.
Thanks, Matt. I appreciate you sharing this article.
Obviously, every point hit the mark - spot on. I would just add one thing: choose your tool intentionally. If I need a visually appealing graphic, GPT image will give me a better result than Claude design. If I care about code quality, I choose Claude. That saves you from the later frustration of wondering why you're not getting what you want.
I love this advice. Choosing tools for the tasks at hand is a great use of critical AI literacy.
As a thinker, I have tried to keep a posture of "trust but verify". I've found it's useful because no one wants to talk philosophy at 3 am.
But, it is a sounding board more than anything. I still keep a period of time where I have time to percolate my ideas. Only then do I start writing.
I will keep those tips in mind as well, to promptmaxx (sorry, couldn't resist) my use going forward.
Ha, I love the term promptmax as it really reveals the Vapidness of the industry.
And, of course, the major fix for all points is the most obvious, don’t use it. Use non AI search tools and verify those sources. Any tool that makes errors isn’t worth using is it? Especially as it creates more time that no using it.
Sometimes it is very helpful though Jason?
Agreed. I tried to post a comment also saying you missed talking about bias. Not all humans are aware of their unconscious biases (most are not) so not all may notice when AI shows its biases
This is a great point, Maha, and AI tools will further serve to exacerbate these biases rather than solve them.
#4 (it is quietly making you worse at your job) is a concern that I saw pop up a great deal in my comments regarding what people are most worried about with AI. There was a huge amount of worry around the cognitive (and skill-related) decline. We need to be deliberate and strategic in our use if we want to preserve the skills we've worked hard to build!
Exactly this. And in order to avoid never-skilling as well.
Your framework is sharp. But the deskilling piece has a genuinely different edge for neurodivergent users that I think is worth naming.
For a mind that has spent decades building compensatory scaffolds for the exact executive functions the world never designed around (outlining, sequencing, task initiation, formal email register), AI isn't replacing a skill that was ever fully theirs. It's replacing an exhausting workaround for a design mismatch.
In that case, "protect the skill you'd lose first" needs recalibration. Sometimes what looks like a lost skill was actually a compensatory tax being paid on the wrong side of the environment. Not always. But often enough that the framework benefits from a second question underneath yours: was this a skill I was building, or a workaround I was maintaining? Different diagnosis, different fix. Loved reading this. 🌿
This is an excellent framing for a position I had not considered. Thank you Alexandra.
One idea I keep coming back to is that AI doesn't remove the need for judgment.
It relocates it.
Instead of spending all our time generating information, we're increasingly deciding when to trust it, when to question it, and when to keep thinking beyond the first plausible answer.
That feels like a fundamentally different cognitive skill than many of us have had to develop before.
This is so well put Karen. It catches all of us in the end...
Thank you, Sam.
That's exactly what fascinates me. The better these tools become, the more interesting the human side of the equation becomes.