Why is AI bad? The real problems with AI in 2026, and the counterarguments

In June 2023, a New York judge sanctioned two lawyers who had filed a brief full of court cases that did not exist. ChatGPT had invented them, complete with plausible names and quotes. Three years later that story no longer looks like a strange one-off. A database of AI hallucination cases kept by researcher Damien Charlotin had logged 1,240 US instances by late July 2026, with 668 of them in 2026 alone.

So why is AI bad? The short answer: AI is not bad in itself, but today's systems make confident mistakes, can copy and amplify human bias, make fraud and deepfakes cheap, raise real questions about jobs, privacy and mental health, and use a lot of energy. Those harms are documented. So are the benefits, which is why the more useful question is when and how AI is bad, and what you can do about it.

We built this explainer from published research, official reports and reputable reporting available as of late October 2026. Where the evidence is thin, we say so.

Short answer: is AI good or bad?

AI is a tool, and like most powerful tools its effects depend on who uses it, for what, and with what safeguards. The main cons of AI are reliability (it makes things up), fairness (it can inherit bias), misuse (scams, deepfakes, non-consensual images), social effects (jobs, over-reliance, mental health) and environmental cost. Against that sit measurable gains in science, medicine and everyday productivity. Our view, explained below, is that most of the harm comes from deploying AI faster than we have learned to check it.

How it works: why AI problems happen in the first place

Most of the AI people use today, such as ChatGPT, Gemini or Claude, is a large language model. In plain terms, it is software trained on huge amounts of text to predict what words should come next. (If you want the longer version, our guide to what AI is covers the basics.) That design explains a surprising number of the problems.

First, a model that predicts plausible text will sometimes produce plausible nonsense. OpenAI's own researchers argued in September 2025 that standard training and testing reward guessing over admitting uncertainty, so models "are encouraged to guess rather than say “I don’t know.”" Think of a student sitting a multiple-choice exam with no penalty for wrong answers. Leaving a question blank guarantees zero, so they always guess. Models learn the same habit.

Second, a model trained on human-written data absorbs human patterns, including unfair ones. It has no built-in sense of which patterns are prejudice and which are fact. Third, the same ability that makes AI useful, producing convincing text, voices and images on demand, is exactly what makes it useful to scammers.

None of this is a bug that one update will fix. It follows from how the technology is built, which is why these AI issues keep coming back in new forms.

Real examples of the negative effects of AI

Confident mistakes with real consequences

The courtroom data is the clearest public record of hallucinations doing damage, because judges write down what went wrong. According to the Maryland State Bar Association's analysis of the Charlotin database, US instances jumped from 11 in 2023 and 37 in 2024 to 524 in 2025, with 97 monetary sanctions that year. Most offenders were people representing themselves, but hundreds were qualified lawyers. If trained professionals get caught out, the rest of us should assume we can be too.

Bias that people then copy

Bias is one of the most cited bad things about AI, and the newest research shows a twist. In a 2025 study, Kyra Wilson, Aylin Caliskan and colleagues asked 528 US participants to screen job candidates alongside an AI that had been set up to favour certain races. With no AI or a neutral one, people picked White and non-White candidates at equal rates. With a strongly biased AI, they followed its lead up to about 90% of the time. Even people who said the AI's advice was unimportant were still swayed.

That matters because the usual safeguard, "a human stays in the loop", assumes the human will catch the error. This study suggests the human may simply absorb it.

Cheaper, more convincing fraud

The FBI's Internet Crime Complaint Center received more than a million complaints in 2025, with reported losses of $20.877 billion. Of those, 22,364 complaints carried an AI-related tag, adding up to $893 million in losses. The breakdown includes voice-cloned "family emergency" calls (more than $5 million) and deepfake job interviews used to get inside company networks (about $13 million). The FBI treats these figures as a floor, since they only count victims who reported the crime and recognised AI was involved.

Harm to vulnerable users

Some of the most serious concerns are about companion chatbots and young people. In January 2026, Google and Character.AI agreed in principle to settle several lawsuits brought by families who alleged the chatbots contributed to teenagers' self-harm and suicides. The settlements contained no admission of liability. Fortune reports that similar cases against OpenAI are ongoing, and that OpenAI denies its products were responsible. Character.AI had already announced in October 2025 that it would bar under-18s from open-ended chats with its characters.

Over-reliance and the "thinking" question

A small MIT Media Lab study, Your Brain on ChatGPT, split 54 participants into three groups to write essays: one with GPT-4o, one with a search engine, one with no tools. The researchers found that "Brain connectivity systematically scaled down with the amount of external support". In the first session, 15 of the 18 ChatGPT users could not correctly quote a line from the essay they had just written.

It is worth being careful here. This is one study, with a small sample and a single task, and it measured engagement during essay writing, not long-term brain damage. What it does support is a common-sense point: if a tool does the thinking, you remember less of it. The internet is also filling up with low-effort machine-written content, which we cover in our piece on AI slop.

Jobs and the environment, in brief

Public worry about work is high. Pew's June 2026 survey found 71% of US adults expect AI to mean fewer jobs over the next 20 years, and 52% are now more concerned than excited about AI in daily life, up from 37% in 2021. The data on actual job losses is more mixed than the headlines suggest; we look at it closely in is AI taking over jobs?

On energy, the International Energy Agency projected in 2025 that data centre electricity use worldwide will reach around 945 terawatt-hours by 2030, more than double, with AI the main driver. Water use for cooling and the location of new data centres add local pressure. Our full breakdown is in is AI bad for the environment?

Benefits and limitations: weighing the cons of AI fairly

A fair answer to "is AI good?" has to include the other column. In 2024, Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for AlphaFold2, which has been used to predict the structure of virtually all of the roughly 200 million known proteins, a problem that used to take years per protein. At a more everyday level, a set of field experiments with 4,867 software developers at Microsoft, Accenture and a Fortune 100 company found that those given an AI coding assistant completed about 26% more tasks, with the biggest gains for less experienced developers.

Here is how we would line up the main AI problems against the evidence and the strongest counterargument for each.

Notice the pattern. The harms with the strongest evidence, such as hallucinations, fraud and bias, are not about AI becoming too clever. They are about AI being trusted too easily. The counterarguments are real, but most of them depend on safeguards that are not yet everywhere.

An analogy we find useful: AI today is like a very fast, very confident new hire who has read everything and checked nothing. Give them a first draft to write and you save hours. Let them sign contracts unsupervised and you will eventually regret it. The problem is rarely the hire. It is the job description.

What it means for you: practical takeaways

So is using AI bad? For most people, no, as long as you treat it as a draft and not a verdict. A few habits do most of the work.

Check anything that matters. Facts, figures, quotes, legal or medical claims and citations should be checked against a primary source. If an answer has no source, ask for one, then open it yourself.

Keep the thinking you want to keep. If you are learning something, write your own first attempt before asking AI, then compare. The MIT study suggests the order matters: participants who started without AI and added it later stayed more engaged.

Be suspicious of urgency. A panicked call from a relative asking for money, or a video of a boss authorising a payment, may be cloned. Agree a family code word and call back on a known number.

Watch who uses it at home. Companion chatbots can feel like friends. For teenagers especially, check age settings and talk about what these apps are and are not.

Mind what you share. Anything you paste into a chatbot may be stored or used under the provider's terms. Do not share passwords, health records or other people's private data.

Our guide on how to use AI covers these habits in more detail.

What happens next

Regulation is arriving, though more slowly than first planned. In June 2026, EU lawmakers reached a provisional agreement to push back key AI Act obligations: rules for high-risk uses such as hiring move to 2 December 2027, and those for AI inside regulated products like medical devices to 2 August 2028. The same deal adds a ban on AI systems that generate child sexual abuse material or non-consensual intimate content from 2 December 2026. For the bigger picture across countries, see our explainer on AI regulation.

What we know is that the documented harms are growing as use grows, and that some of them, like courtroom hallucinations, are rising sharply. What we are guessing about is the long run: whether over-reliance changes how people learn, how far jobs shift, and whether better training methods can cut hallucinations substantially. OpenAI's own research points to one fix, rewarding models for saying "I don't know", but whether the industry adopts it widely is an open question.

Where do we land? AI is not bad in the way a poison is bad. It is risky in the way a car is risky: very useful, occasionally deadly, and much safer with rules, training and attention. The reasons AI is bad today are mostly reasons we are using it carelessly. That is a more hopeful conclusion than it sounds, because carelessness is something we can fix.

FAQ

Is AI good?

AI can be very good at specific tasks. It helped predict the structures of about 200 million proteins and, in field experiments, helped developers complete roughly 26% more tasks. It is less good at being reliably accurate or fair on its own. Whether it is good overall depends on the task, the safeguards and how carefully people check its output.

How is AI bad?

AI is bad mainly in five ways: it makes confident factual errors, it can reproduce bias, it makes scams and deepfakes cheaper, it can harm vulnerable users such as teenagers using companion chatbots, and it uses large amounts of electricity and water. Many of these harms come from people trusting its output without checking.

What is bad about AI?

The most documented problems are hallucinations, such as invented court cases, bias in decisions like hiring, and fraud. The FBI linked $893 million in reported 2025 US losses to AI-related scams. Concerns about jobs, privacy and over-reliance are widespread, but the evidence on their long-term effects is still developing.

Is AI harmful?

It can be. Lawsuits have alleged that chatbots contributed to teenagers' self-harm, voice-cloning scams have cost families money, and biased recommendations can sway human decisions. But AI is not harmful by default. Most harm depends on how it is designed, deployed and supervised, which is why regulation and safeguards matter.

Why is using AI bad?

Using AI is not bad in itself. It becomes a problem when you rely on it for facts without checking, let it do thinking you need to learn, share private data with it, or use it to deceive people. Used as a starting point that you review and edit, it is a reasonable tool for most everyday tasks.