Is AI Bad? The Real Harms and Benefits, With Sources

In brief

AI is bad in measurable ways and good in measurable ways. GPT-3.5 invented 55% of its citations in a 2023 study, and the IEA says data centres used 485 TWh of electricity in 2025. A study of 5,179 support agents found AI raised output 14%. Judge each use on its own evidence.

What are the pros and cons of AI?

AI has real harms and real benefits, and the evidence for each is specific. The table lists one sourced finding per area, read from the source on 22 November 2026.

An old brass balance scale on a black plinth, one pan holding a small rainbow glass Gemini star and the other a terracotta Claude starburst, with a brushed silver OpenAI knot resting on the plinth beside it, in a charcoal studio under white rim light.
AreaPro or conWhat the evidence saysSource
Facts and sourcesConGPT-3.5 fabricated 55% of 222 cited works, GPT-4 fabricated 18% of 414Scientific Reports, 2023
FairnessConMany face recognition algorithms had false positive rates 10 to 100 times higher for some groupsNIST, 2019
JobsBoth"almost 40 percent of global employment is exposed to AI"; about half of exposed jobs in rich countries may benefitIMF, January 2024
Work outputPro14% more issues resolved per hour, 34% for novices, across 5,179 agentsNBER working paper 31161
ScamsConCriminals use generative AI to commit fraud "on a larger scale" with fake images, cloned voices and videoFBI, December 2024
EnergyConData centres used 485 TWh in 2025 and are projected at 950 TWh in 2030IEA, 2026

No single row answers the question, which is why "is AI bad" has no one word answer. The rest of this page takes each row in turn, says what the number does and doesn't show, and ends with questions you can put to an AI yourself.

Does AI make things up?

Yes. Language models state invented facts in the same confident voice as true ones, and the best measured case is citations. In 2023 Walters and Wilder had GPT-3.5 and GPT-4 write 42 short papers each and checked the 636 citations in them.

In a 2023 test, 55% of GPT-3.5's citations were invented. GPT-4 invented 18%
GPT-3.5222 citations55% inventedGPT-4414 citations18% inventedone square is about 2 percent of that model's citations

Real case: a student asks a chatbot for sources and pastes the list into a bibliography. Some of the works don't exist. Among the real ones, 43% of GPT-3.5's and 24% of GPT-4's carried at least one substantive error, and fewer than 10% came with a link, so a link you can open is the quick test.

That study tested models from 2023, so it says little about today's models, and it doesn't report whether either had web access. Our pages on what an AI hallucination is and hallucination statistics cover the newer numbers, and how to fact check AI answers gives the two minute routine.

Is AI biased?

In several well tested cases, yes. The clearest public measurement is the US National Institute of Standards and Technology study of face recognition, released in December 2019, which tested 189 algorithms from 99 developers on 18.27 million images of 8.49 million people.

NIST wrote that it found "empirical evidence for the existence of demographic differentials in the majority of the face recognition algorithms". For one to one matching, false positive rates for Asian and African American faces often ran 10 to 100 times those for Caucasian faces, depending on the algorithm. Not every algorithm showed it, and the most equitable ones were also among the most accurate.

Real case: a face match used to flag a suspect or unlock a phone. A false positive means the wrong person is matched. That is why the harm lands unevenly.

The University of California page that ranks third for this query says facial recognition is less accurate for women and people of color but cites no study for it. NIST is the study. It tested face matching, not chatbots, so it doesn't prove a chatbot is biased. The mechanism is the same: a system trained on unrepresentative data does worse on the people it saw least.

Is AI bad for jobs?

Nobody can say yet, and the best sources measure exposure rather than layoffs. The IMF's managing director wrote in January 2024 that "almost 40 percent of global employment is exposed to AI." For advanced economies she put it near 60 percent, with roughly half of exposed jobs likely to benefit from AI integration and the other half facing lower labor demand.

Exposed means AI can do part of the work. It doesn't mean the job ends. AI job loss statistics separates forecasts from counted layoffs, and that is the page to read before you believe a headline number.

Real case: the benefit side is measured too. Brynjolfsson, Li and Raymond studied 5,179 customer support agents given an AI assistant. Issues resolved per hour rose 14% on average and 34% for novice and low skilled workers, while experienced workers saw little effect. One company's call centre isn't the whole economy, but it is a named study with a sample size.

Is AI bad for the environment?

It is a growing load on electricity, though still a small share of the total. The International Energy Agency counts data centres, not AI alone: they used about 415 TWh in 2024, around 1.5% of global electricity, and 485 TWh in 2025.

Its 2026 report projects the figure "roughly doubling from 485 TWh in 2025 to 950 TWh in 2030", about 3% of global demand. It says consumption from AI focused data centres grew 50% in 2025 and is projected to triple by 2030. These are forecasts, and the IEA's 2025 report ran a higher and a lower case out to 2035 around its base case.

Two limits matter. Data centres also run email, video and search, so the 485 TWh isn't all AI. And the pages I opened for this one give electricity, not a water figure, so I'm not quoting one. AI data center energy statistics goes deeper on per prompt costs.

Is AI dangerous?

Some uses are, and the dangers people argue about fall in three groups: wrong answers acted on, misuse by people, and long term risk from far more capable systems. The first is measured above. The second and third are mostly judgment, not data.

Virginia Tech's engineering magazine, the page that ranks first for this query, collects five faculty views. They raise bias, a large carbon footprint, job security, reduced critical thinking from overreliance, and AI used as a tool of propaganda. Those are credible concerns from named experts, but the piece gives no figures for any of them, which is why this page adds the numbers above.

Privacy is its own risk with its own answer: is AI safe to use covers where your messages go and the settings to check. For the long term worries, will AI take over the world sets out what experts argue and where they disagree. Rules and responsibility are in AI ethics.

Can AI be used for scams, deepfakes and misinformation?

Yes, and the FBI says so in plain terms. Its Internet Crime Complaint Center warned the public that criminals exploit generative AI "to commit fraud on a larger scale which increases the believability of their schemes," and that it "reduces the time and effort criminals must expend to deceive their targets."

The warning lists concrete cases: fake profile photos for romance and investment fraud, forged identity documents, short audio clips of the voice of a loved one used to demand money, and real time video chats with someone posing as an executive or official. The FBI gave no victim count in that notice, so this is a documented pattern, not a measured rate.

Real case: a call arrives in a relative's voice asking for urgent money. The fix is dull and works: hang up and call back on a number you already have. This is misuse by people, so it is a different problem from the model errors above, and the Virginia Tech and University of California pages that rank for this query raise the same worry about fake content and propaganda without giving a source for it.

Is there anything good about AI beyond work?

Yes, in health and accessibility, though the strongest evidence is narrower than the hype. The US Food and Drug Administration keeps a public list of AI enabled medical devices it has authorized, and the list carries entries dated into June 2026. Authorization means a device met the FDA's bar for its stated use. It doesn't mean AI beats doctors in general.

Virginia Tech faculty point to assistive robot arms and wheelchairs, cancer detection and AI that helps professionals instead of replacing them. Those are named expert views, not measurements. Treat them as reasons the benefits side is real and ask for the study behind any specific claim, the same as for a harm.

What questions should you ask an AI about whether AI is bad?

Ask questions that force sources, because an AI answering "is AI bad" without them gives you an average of opinions. Paste any of these into Claude, ChatGPT, Gemini or Whizi and fill the brackets.

  1. Get both sides: "Give the strongest evidence that [claim about AI] is true and the strongest that it is false, three sentences each, with a link to a primary source for each side. Say which side the evidence favors and how sure you are."
  1. Audit a number: "A site says [statistic about AI]. Find the original study or report, tell me the date, the sample and the model tested, and say whether the site stated it correctly."
  1. Check my job: "I work as [job]. List the tasks in it that AI tools can do now, the tasks they can't, and one source for each claim. Don't guess about the future."
  1. Cost of one task: "How much electricity does [task, such as one chatbot answer] use? Use only the company's own report or a named measurement, give the date, and say what it leaves out."
  1. Stress test an answer: "Here is an answer an AI gave me: [paste]. Which claims could be invented? Mark each as checkable or not, and tell me where to check."
  1. Explain it at my level: "Explain why AI is [accused of bias or making things up] to a [grade] student, using one real case with a named source, then ask me one question to check I got it."

What mistakes make people wrong about AI?

Five mistakes cause most bad takes on this topic, and each has a fix.

  1. Treating one old test as today's result. The fix: check the model and the date. The citation study above tested 2023 models.
  2. Reading exposure as job loss. The fix: look for counted layoffs, not only forecasts.
  3. Quoting a total with no baseline. The fix: ask what share of the whole it is. 485 TWh sounds huge until you see the IEA puts it near a few percent of global electricity.
  4. Trusting an AI summary of the debate. The fix: use prompt 2 and read the original.
  5. Judging "AI" as one thing. A face matcher, a chatbot and a call centre assistant fail in different ways, and the studies above tested one each. The fix: name the use and the system before you decide.

A good test of any claim you meet is whether it names a study, a date and a number. The two pages at the top of Google for this query name experts and give almost no numbers, so you can do better by asking for all three every time. Most arguments about AI go away once everyone is looking at the same measurement.

Where can you try the models and judge for yourself?

Whizi puts Claude, GPT and Gemini in one app, so you can ask the same question of each and read where they disagree. Web search is on by default in each chat, which makes it easier to ask for sources.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, no card. Pro carries Claude, GPT and Gemini in one app, its first week is $1.99, and side by side Compare runs two models on one prompt on Pro and Powerhouse. See Whizi pricing for what each plan holds.

What can this page not tell you?

It can't tell you whether AI is a net good for society, because that depends on values as well as data. It reports what primary sources measured, and each has limits.

The citation study is from 2023 and covered two models. The NIST study covered face recognition, not chatbots. The IMF number is exposure, not layoffs. The NBER study is one firm's support staff. The IEA figures are projections. A balanced answer holds all of those at once and updates when new evidence arrives.

Workflow checklist
  • Open the original study before you repeat a statistic about AI
  • Check the model and the date behind any claim about accuracy
  • Separate jobs exposed to AI from jobs actually lost
  • Ask what share of the whole a big energy number really is
  • Ask two different models and open every source they give
Common questions

Frequently asked questions

Is AI bad for the environment?

AI adds to electricity demand, though data centres are still a small share of the total. The IEA counts 485 TWh for all data centres in 2025 and projects 950 TWh in 2030, about 3% of global electricity. That figure includes email, video and search, so not all of it is AI.

Is AI bad for your brain?

No study we opened settles it. Virginia Tech computer scientist Eugenia Rho argues in the university's engineering magazine that depending on AI could reduce critical thinking, which is a stated concern, not a measured result. Using AI to check your own work instead of replacing it is the safer habit.

Is AI dangerous?

Some uses are. Wrong answers acted on are the best measured danger: GPT-3.5 invented 55% of its citations in a 2023 study. Misuse and long term risk are debated rather than measured. Check any high stakes answer with a source you can open and a qualified person.

Will AI take my job?

Nobody knows yet. The IMF says almost 40 percent of global employment is exposed to AI, but exposed isn't lost, and about half of exposed jobs in advanced economies may benefit. Our AI job loss statistics page separates forecasts from counted layoffs.