Artificial intelligence is a net gain wherever its output is cheap to check and a net risk wherever errors are expensive and hard to spot. The pros are concrete: it cuts the time of pattern-heavy work such as transcription, drafting, image sorting and code review by roughly a third to a half in controlled studies, it has produced real scientific advances such as protein-structure prediction, and it makes accessibility tools like live captions nearly free. The cons are equally concrete: it states wrong answers with the same confidence as right ones, it inherits bias from training data, it displaces specific job categories faster than it creates replacements, it consumes a fast-growing share of electricity, and it hands fraudsters cheap voice and video impersonation. The sections below put numbers on each.

Where does AI measurably help?

The strongest evidence is for mid-skill knowledge work. A randomized experiment published in Science found that professionals using ChatGPT for writing tasks finished 40 percent faster with higher-rated output, and the gain was largest for the weakest writers. A field study of 5,000 customer-support agents, published by the National Bureau of Economic Research, found a 14 percent productivity increase on average and 34 percent for novice agents.

In science and medicine the gains are harder to dismiss as hype. AlphaFold’s structure predictions, published in Nature, solved a 50-year-old problem and earned a share of the 2024 Nobel Prize in Chemistry. The U.S. Food and Drug Administration’s list of authorized AI-enabled medical devices passed 1,000 entries in 2025, most of them in radiology, where the software flags findings for a human to confirm rather than diagnosing on its own.

Accessibility is the quiet success. Automatic captions, screen-reader image descriptions and real-time translation are built into phones and video platforms at no marginal cost. Ten years ago each was a specialist service billed by the minute.

What does it cost, and who pays?

Energy. The International Energy Agency’s Energy and AI report puts data-centre consumption at about 415 TWh in 2024, roughly 1.5 percent of global electricity, and projects it to more than double by 2030 with AI the main driver. That load lands on specific grids, which is why utilities in Virginia, Ireland and the Netherlands have paused new connections.

Jobs. The World Economic Forum’s Future of Jobs Report 2025 estimates 170 million roles created and 92 million displaced by 2030. The net figure hides the problem. The displaced roles are concentrated in clerical, data-entry, customer-service and entry-level creative work, and the people in them are not the ones hired for the new roles.

Concentration. Training a frontier model now costs hundreds of millions of dollars in compute alone, so a handful of companies set the terms for everyone downstream. Small firms do not build models. They rent them, and pricing, usage policy and model retirement are outside their control.

Why are bias and errors the hardest cons to manage?

Bias is documented, not hypothetical. The National Institute of Standards and Technology’s face-recognition demographic study found false-positive rates for some algorithms 10 to 100 times higher for Asian and Black faces than for white faces. The earlier Gender Shades study found commercial gender classifiers erred on up to 34.7 percent of darker-skinned women versus under 1 percent of lighter-skinned men. Systems trained on hiring, lending or policing data reproduce the historical pattern in that data unless someone tests for it.

Errors are a different problem. Language models generate plausible text, and plausibility is not accuracy. A fabricated citation, a wrong dosage or an invented legal precedent reads exactly like a correct one. The practical rule is that AI output needs a reviewer who could have produced it themselves. Where no such reviewer exists, the error rate is unknown.

How does AI change security?

Defensively, AI is now standard in anomaly detection and malware classification. Offensively, the same models write convincing phishing in any language and clone a voice from a few seconds of audio. The OWASP Top 10 for LLM Applications catalogues a new class of attacks, led by prompt injection, where an attacker hides instructions in content the model reads.

Deepfakes have moved from novelty to fraud tool. The realistic defence is not detection, which is losing the arms race, but provenance. The Coalition for Content Provenance and Authenticity standard attaches signed metadata at capture and through each edit, and it is now built into cameras from Leica, Sony and Nikon and into Adobe’s editing tools.

What has changed in the rules since early 2025?

• The European Union’s AI Act is in force. Bans on practices such as social scoring applied from February 2025, obligations for general-purpose model providers from August 2025, and most high-risk-system rules are scheduled for August 2026, though the Commission has proposed delaying some of them.
• ISO/IEC 42001 is the first certifiable AI management-system standard and is becoming a procurement requirement in regulated sectors.
• The NIST AI Risk Management Framework and its generative-AI profile are the de facto U.S. baseline in the absence of federal legislation. State laws, led by Colorado, fill the gap unevenly.
• The U.S. Copyright Office has confirmed that purely AI-generated material is not copyrightable. Only human-authored selection, arrangement or modification earns protection.

What does AI actually do inside a video production company?

We use it daily, and the pattern matches the research: large gains on checkable tasks, little on judgment. Transcribing a 60-minute interview used to cost 4 to 6 hours of an editor’s time or about $1 to $1.50 per minute outsourced. It now takes minutes, and text-based editing lets us assemble a first rough cut from the transcript before touching a timeline. Captioning, rotoscoping, audio clean-up and frame extension for vertical crops are similar wins.

What goes wrong is specific. Transcription models mangle client brand names and proper nouns, so every caption file still gets a human pass. Generative video tools produce 5 to 10 second clips, and in our experience it takes 10 to 30 generations to get one usable insert, with hands, text and continuity the usual failures. That makes AI footage viable for 10 to 20 percent of B-roll in an explainer and unviable for a hero shot where a client’s product has to be exactly right.

The trade-offs are legal as much as technical. Cloning a presenter’s voice for a pickup line requires written consent, and for union talent the 2023 SAG-AFTRA agreement requires consent and compensation for any digital replica. Several stock libraries refuse AI-generated submissions, and because AI output cannot be copyrighted, a client who wants exclusive ownership of their spot cannot get it for the AI-generated parts. We disclose every AI element in the deliverable notes, because the client’s legal team will ask eventually and the answer should already be on file.

How should you decide whether to use it?

Start by listing the tasks your team does that a competent reviewer can verify in a fraction of the time it takes to do them. Pilot AI there, measure the time saved and the error rate for a month, and write down who reviews what. Then map any use touching customers, hiring or personal data against the NIST framework or the EU AI Act’s risk tiers before scaling it, because those are the uses where the cons are not theoretical.