The Promise and Risk of AI: What It Delivers, What It Costs
What does AI deliver, and what does it cost?
The promise of AI is speed on narrow, checkable work: drafting, summarizing, coding, translating, and finding patterns in more data than a team could read. Its risk comes from the same mechanism: a model that predicts likely text states wrong answers as confidently as right ones and cannot tell an instruction from data, so it can be fooled and hijacked. Across 84 scored episodes of Promise & Risk of AI, 80 came out balanced, four leaned risk and none leaned promise. Inside the balanced band, 48 scored below zero and 22 above it, and most of that risk sat in AI that organizations had already deployed.
On 31 March 2026, the first episode of this channel promised one thing: a consistent way to weigh any AI claim. Every episode since has ended on a verdict, a score from −100 (all risk) to +100 (all promise). 84 episodes carry one. The average is −8.
That number is the answer to the question in the title, and the rest of this page is the working. Three questions come up more than any others when people search for the promise and risk of AI. Here is what the episodes found.
What is the promise of AI?
The promise of AI is real, and narrower than the marketing. Episode 12 gave it a three-word test: narrow, measured, monitored. Where a task passes that test, AI does work no team could match on speed or cost.
- Writing and analysis. Drafting, summarizing, brainstorming and code completion save real hours when a person edits the result (Episode 3). A four-part prompt (role, context, task, format) closes most of the gap between a mediocre answer and a good one. Episode 8, which teaches it, scored +30.
- Medicine. Narrow, FDA-cleared models read a retinal scan or flag a stroke on a CT. Each was tested against real patients and cleared for one task (Episode 12).
- Money. Banks have run machine-learning fraud detection for more than 20 years, watching transaction volumes no review team could (Episode 11).
- Security. The machine speed that arms attackers arms defenders too, and the playbook to use it already exists (Episode 18).
- Work. AI automates tasks inside a job, and most of the job stays. When ATMs arrived, the number of bank tellers went up (Episode 16).
Outside those conditions, the same tool is a fluent guesser, and the bill arrives later.
What are the five biggest risks of AI?
Five risks came up in episode after episode. Most of them follow from how the technology works, so a better model does not make them go away.
1. Confident fabrication
A large language model predicts the next word instead of retrieving a fact. Wrong answers arrive in the same confident voice as right ones, which is how a general chatbot can hand someone a wrong methotrexate dose. Three in five adults have asked a chatbot a medical question (Episode 12).
2. Prompt injection
Instructions and data reach a model on one stream of tokens, and there is no second channel to keep them apart. Prompt injection has topped the OWASP Top 10 for LLM applications in every edition since 2023 (Episode 24). Self-hosting does not fix it: a model on your own machine is exactly as injectable as the same model behind an API (Episode 25).
3. Over-permissioned AI agents
The exposure most organizations carry is an AI agent they set up themselves, with more access than anyone reviewed (Episode 18). Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Of the thousands of vendors selling agentic AI, it puts roughly 130 in the real category (Episode 20).
4. Deepfake fraud
In 2024, a finance employee at Arup in Hong Kong approved $25 million on a video call where every other face was a deepfake (Episode 10). In a test of 2,000 people who were warned to look for fakes, one in a thousand got every answer right (Episode 22). Americans lost $12.5 billion to fraud in 2024, up 25% in a year while the number of scams held flat (Episode 11).
5. Data that outlives your settings
Training, retention and memory each have their own switch, and the training opt-out most people find touches only the first (Episode 21). A court can override a setting you used correctly: in May 2025, a federal magistrate ordered OpenAI to preserve output logs, including conversations users had deleted (Episode 21).
The cheapest control on this list takes about 20 minutes, and it limits risks two and three. List every action your AI agent can take without approval, mark the ones an attacker would want, and put those behind a person (Episode 24).
Is AI a threat or a promise?
Both, and the split can be counted. Of 84 scored episodes, 80 came out balanced on the meter, between −30 and +30. Four leaned risk and none leaned promise: no episode has scored above +30. Inside the balanced band, 48 scored below zero and 22 above it, and the pattern sits in what the AI touches. Episodes about using AI to do your own work better mostly score above zero. Episodes about AI given access to money, identity or systems score below it.
After 25 years in cybersecurity, I’ve watched this arc with cloud and mobile: the capability ships first, the controls follow years later, and the losses land in the gap. AI is following the same arc. Faster.
On 16 June 2026, in the episode on medical AI, I put the distinction in four words: “Same label. Completely different risk.” It holds well beyond medicine. The table below shows where.
| Area | What AI delivers | What it costs | Verdict |
|---|---|---|---|
| Prompting | A four-part prompt (role, context, task, format) gets far better answers from the same model | A prompt that works on one model can fail silently on another | +30 balanced |
| How chatbots work | Drafting, summarizing, brainstorming and coding on demand | Fabricated answers, delivered as confidently as true ones | +10 balanced |
| Privacy settings | Training, retention and memory each have a switch that works | Opt-outs are not retroactive, and memory keeps writing after training is off | +10 balanced |
| Jobs | AI automates tasks inside a role, and most of the role stays | Workers aged 22 to 25 in the most exposed jobs are down about 13% since 2022 | −5 balanced |
| Money and fraud | Banks have run machine-learning fraud detection for more than 20 years | The same engine powers the scam; Americans lost $12.5 billion to fraud in 2024 | −10 balanced |
| EU AI Act | The binding rules are specific: transparency, documentation, human oversight | The headline delay moved the high-risk rules; the rest kept their dates | −10 balanced |
| Medicine | Narrow FDA-cleared tools, each validated for one task on real patients | General chatbots, cleared for nothing, give confident wrong doses | −20 balanced |
| Cybersecurity | Defenders get the same machine speed attackers do | The biggest exposure is the over-permissioned agent you set up yourself | −25 balanced |
| Prompt injection | A 20-minute tool-permission audit limits what a hijacked agent can do | Top of the OWASP list for LLM applications since 2023, with no fix | −25 balanced |
| Deepfakes | A callback to a number already in your phone still works | One person in a thousand spotted every fake in a test | −30 balanced |
| AI agents | A real agent re-plans inside guardrails you set | Much of what is sold as an agent is relabeled automation on a consumption meter | −30 balanced |
| Self-hosting | You control the context, the retrieval corpus, egress and the tools | The same injection risk, without the vendor’s defense layers | −35 risk |
The promise is in what you do with AI. The risk is in what you let it do.
Questions people ask
- What is the promise of AI?
-
AI’s promise is speed on narrow, checkable work: drafting, summarizing, coding, translating, and finding patterns in more data than a team could read. In medicine, narrow FDA-cleared tools that read scans are each validated for one task, and banks have used machine learning against fraud for more than 20 years. The promise holds when the task is narrow, the result is measured, and a person monitors it.
Episode 8: the prompt framework → - Is AI a threat or a promise?
-
Both, and the split can be counted. Across 84 scored episodes of Promise & Risk of AI, 80 came out balanced on the meter (between −30 and +30), four leaned risk and none leaned promise. Inside the balanced band, 48 scored below zero and 22 above it. AI used to improve your own work tends to score above zero; AI given access to money, identity or systems tends to score below.
Episode 19: the AI you already have → - What are 5 risks of AI?
-
One, confident fabrication: language models predict likely words and state wrong answers as confidently as right ones. Two, prompt injection: instructions and data share one channel, so a document or web page can hijack the model. Three, AI agents with more access than anyone reviewed. Four, deepfake fraud, such as the $25 million Arup video call. Five, data that outlives your settings, because training, retention and memory are controlled separately.
Episode 24: prompt injection → - Which jobs will survive AI?
-
Most jobs survive in changed form, because AI automates tasks inside a job and most roles keep the tasks it cannot do. When ATMs arrived, the number of bank tellers went up as the role moved toward work the machine could not do. The early damage lands on entry-level roles: workers aged 22 to 25 in the most AI-exposed jobs are down about 13% since 2022, while workers over 30 are roughly flat.
Episode 16: tasks, not jobs → - How does Promise & Risk of AI score an episode?
-
Every episode ends on a verdict from −100 (all risk) to +100 (all promise), shown on the Promise & Risk meter. The score weighs what the technology delivers when used well against what goes wrong in real deployments. A score of 0 means balanced.