Most AI risks are not glitches. They are design choices.
Every risk below has a cause that traces back to a decision someone made — about
what data to collect, what to optimise for, or what to leave invisible. Each row
opens to show three layers: what the risk is,
why it exists, and what you can actually do about it.
01
Bias
Models inherit the assumptions of whoever chose the training data.
What
A model trained on past loan decisions, hiring records, or arrest data
learns the patterns inside that data — including the patterns no one
wanted it to learn. The result is a system that quietly recreates old
unfairness in a new, more confident-sounding form.
Why
Models do not understand fairness. They optimise for whatever metric
they were given, on whatever data they were given. If the data reflects
decades of human bias, the model will do the same — only faster, at
scale, and with the appearance of mathematical objectivity.
What you can do
When a system makes a decision about you, ask what data it was trained
on and what outcome it was optimised for. "The algorithm decided" is
never a real answer — someone chose the training set, the goal, and
the threshold.
02
Privacy
The price of free services is often a copy of you that never expires.
What
Every prompt, every photo, every voice command is potential training
data. Once it has been absorbed into a model, it cannot be reliably
removed — the data is gone, but its influence on the model's behaviour
remains.
Why
Modern AI requires enormous amounts of data, and the cheapest data
source is the people using the product. The business model rewards
collecting more, not less, and "delete my data" laws are difficult
to enforce on systems whose memory is statistical rather than literal.
What you can do
Read the data setting on any AI tool you use, and turn off training
data collection where the option exists. Treat free chatbots like
public squares: assume what you say can be remembered.
03
Misinformation
Convincing falsehoods are now cheaper to produce than the truth.
What
Generative AI can fabricate articles, voices, photographs, and video
of events that never happened. The technology that lets a podcaster
clean up audio is the same technology that lets someone fake a
confession.
Why
For most of human history, producing convincing fake media required
time, skill, or money. AI removes all three. The constraint that
quietly protected public trust — that lying well was hard — is gone.
What you can do
Stop trusting media because it looks real. Start trusting it because
of who is willing to put their name on it. Provenance — knowing where
something came from — is now more important than how convincing it
seems.
04
Concentration of power
Training the most capable models requires resources only a few entities have.
What
A handful of companies control the largest AI models, the data
centres they run on, and the rules for who is allowed to use them.
Increasingly, what AI can and cannot say is decided by a small number
of private boardrooms.
Why
Training a frontier model costs hundreds of millions of dollars in
compute alone. That price tag is itself a barrier — most universities,
most governments, and almost every individual are spectators rather
than participants in how the technology develops.
What you can do
Support open-weight models, transparency standards, and competition
policy that keeps the field plural. A monoculture of AI providers is
a fragile thing — for users, for innovation, and for democracy.
05
Accountability gap
When an AI system causes harm, no one is sure who is responsible.
What
A flawed AI denies a benefit, misdiagnoses a condition, or rejects a
CV. Who is liable? The developer who built the model? The company
that deployed it? The data provider? The user who clicked accept on
the terms of service?
Why
AI systems are built by long supply chains. Each link in the chain
disclaims responsibility for what the others do. Existing law was
written for products with clear authors — software that learns from
data does not fit neatly into that frame.
What you can do
Insist on a human in the loop for decisions that materially affect
you — credit, medicine, justice, employment. The right to an
explanation, and the right of appeal to a person, are worth fighting
for now while the rules are still being written.
06
Dependence
Skills we stop using are skills we eventually lose.
What
Calculators changed how we do arithmetic. GPS changed how we navigate.
Generative AI is changing how we write, plan, and think. The question
is not whether we adapt, but which capacities we are willing to
outsource — and which we should defend.
Why
The brain is efficient: practised skills strengthen, unused skills
fade. When a tool is always available, the underlying ability often
atrophies. This is fine for arithmetic. It is more concerning for
judgment, writing, and reasoning.
What you can do
Use AI as a collaborator, not a substitute. Draft your own thinking
first, then let the model challenge it. The goal is to leave a
session with sharper reasoning than you arrived with — not a longer
output you didn't write.
None of these risks are inevitable. Every one of them is a choice —
made by someone, somewhere, that can be questioned, contested, or unmade.
Awareness is the precondition for choosing differently.
Keep reading
These risks have a counterweight — see real projects making a difference in
AI for Good,
or trace the risks back to their technical roots in
Understanding AI.