AI ethics, ticked off honestly.
Tickd tracks every major AI platform and is owned or funded by none of them. That independence is the whole point — and it comes with an obligation. If we're going to help you fix these tools, we should also be straight with you about the harder questions the tools raise.
So this is not a troubleshooting section. It's where we look at alignment, bias, privacy, copyright, jobs, agents, energy and regulation properly — the arguments on each side, who makes them, and where the evidence is thin.
We're not here to sell you on AI, and we're not here to frighten you about it. On genuinely contested questions — how severe alignment risk is, which regulatory model works, how fast any of this arrives — we lay out the credible positions and leave the conclusion to you. Same unbiased-by-design approach as everything else on Tickd.
The eleven questions
Every serious debate in AI ethics fits somewhere in here. Start anywhere.
- 01
The Alignment Problem
Getting a system to do what we actually meant, not just what we literally asked for. Some researchers treat it as a hard engineering problem that will yield to better tooling; others see it as the defining safety question of the century. Both camps have serious people in them.
- 02
Bias & Fairness
Models learn from human output, and human output carries our patterns — including the ugly ones. The debate is less about whether bias exists and more about who gets to decide what a corrected model should say instead.
- 03
Misinformation & Deepfakes
Synthetic video, cloned voices and convincing fake documents are now cheap. Detection tools help, but the arms race favours the generators — so media literacy is doing more work than technology right now.
- 04
Privacy & Data Use
What happens to what you type? It depends on the product, the plan and the settings — and the defaults differ wildly between platforms. Knowing the difference between chat history, training data and retention logs matters.
- 05
Copyright & Creative Rights
Courts across several jurisdictions are still deciding whether training on copyrighted work is fair use, and who — if anyone — owns the output. Image generators sit at the centre of the fight.
- 06
Jobs & Economic Disruption
Forecasts range from mass displacement to modest productivity gains. The measured studies so far show task-level change more than wholesale job deletion — but they are early, narrow and contested.
- 07
Agentic AI Risk
When a model stops answering and starts acting — booking, buying, writing to your files — mistakes get expensive. The safety questions shift from what it says to what permissions it holds.
- 08
The Black Box Problem
Nobody, including the labs, can fully explain why a large model produced a specific answer. Interpretability research is making real progress, and is still nowhere near a full account.
- 09
Concentration of Power
Frontier training runs need capital and compute that only a handful of organisations have. Whether open-weight models meaningfully offset that is one of the sharpest disagreements in the field.
- 10
Environmental Cost
Training and serving models uses electricity and water. Estimates vary by orders of magnitude depending on assumptions, and disclosure from operators remains patchy.
Deep-dive writing on this one is in the works.
- 11
Regulation Landscape
The EU regulates by risk tier, the US mostly by sector and state, China by content and licensing. Plain-English overview only — nothing here is legal advice.
Latest ethics writing
New pieces land here as we publish them.
Why You Shouldn't Use LLMs to Draft Your App's Security Vulnerability Disclosures (And the Ethical Way to Handle Breaches)
When a security incident strikes, the pressure to publish an update is immense. Here is why letting an LLM write your security disclosures or CVE write-ups is incredibly risky, and how to co-author them ethically.
Why You Shouldn't Use LLMs to Moderate Your Online Community (And the Ethical Way to Keep Spaces Safe)
Slapping an LLM API onto your Discord server or community forum to auto-moderate posts seems like a developer's dream. Here is why automated algorithmic justice backfires, and how to build ethical safety nets instead.
Why You Shouldn't Use LLMs to Simulate User Personas (And the Ethical Way to Do User Research)
Generating 'synthetic users' in Claude or Gemini to test your product features sounds like a dream. In reality, it is an echo chamber that excludes real-world human experience.
Why You Shouldn't Use LLMs to Grade Developer Code Assessments (And the Ethical Way to Review Technical Tests)
Hiring is exhausting, and letting an AI grade those pile-high coding tests is incredibly tempting. Here is why automated LLM grading is a shortcut to hiring the wrong people—and how to build an ethical, human-first review pipeline instead.
Why You Shouldn't Use LLMs to Write Your Team's Incident Post-Mortems (And the Ethical Way to Document Outages)
When production goes down, writing the retro is a chore. But outsourcing your incident reports to an LLM sanitises critical human errors and destroys engineering culture. Here is how to keep retrospectives human.
Why You Shouldn't Use LLMs to Ghostwrite Your Tech Blog (And the Ethical Way to Co-Author with AI)
It is tempting to outsource your developer blog to an LLM. But the moment you publish plastic, AI-generated prose, you lose the trust of the engineering community. Here is how to use AI as an editor, not a ghostwriter.
Key terms
The vocabulary these debates run on. Full definitions live in the Tickd glossary.
- Alignment
- Making a model's behaviour match human intent, not just instructions.
- RLHF
- Reinforcement learning from human feedback — how preferences get baked in.
- Interpretability
- Research into why a model produced a given output.
- AGI
- A contested label for systems matching humans across most cognitive work.
- Hallucination
- Confident output that isn't true.
- Guardrails
- The refusal and safety layer sitting on top of a base model.
- Fine-tuning
- Further training that shapes tone, behaviour and limits.
- Agent
- A model given tools and permission to take actions.
- Training data
- The corpus a model learned from — and the root of most bias debates.
- Open weights
- Models whose parameters are published for anyone to run.