Generative AI works.
That is the problem.
Dr Cong Ngo · Postdoctoral Research Fellow · Monash University
These systems do what they promise. They draft, summarise, translate and screen faster than any team you could hire. What none of them can do is tell you when they are wrong — the answer arrives fluent and confident either way, carrying no record of where it came from and nobody answerable when it fails.
So the bottleneck has moved. Producing work is now nearly free; checking it costs exactly what it always did. In research, teaching and public policy that gap is already being paid for quietly — in citations nobody verified, in sound evidence that never surfaced, in decisions no one can reconstruct a year later. Closing it is practical work, and it is what I do: the research to establish what is actually true, and the tools that make the careful path the easy one.
Where the conversation sits
Two things are true at the same time.
It is not useful to argue that generative AI does not work. It does, and the gains in some tasks are real and measurable. It is equally not useful to pretend that capability and trustworthiness arrived together. They did not.
What genuinely arrived
- Drafting, restructuring and translation at a cost close to zero per attempt.
- First-pass triage of large document sets that no team could read in the time available.
- Pattern-matching across text that would previously have taken a dedicated analyst.
- A genuine lowering of the technical barrier: people who could not build software can now build tools.
- Accessibility gains — for non-native writers, for people with disabilities, for under-resourced institutions.
What did not arrive with it
- Any reliable signal of when an output is wrong. Fluency is constant; accuracy is not.
- Provenance. The sources behind an answer are, by default, unrecoverable.
- A shared convention for declaring how AI was used in a piece of work.
- Verification capacity. Output volume rose; the number of qualified people checking it did not.
- Clear accountability. When an AI-assisted claim fails, the chain of responsibility is usually undocumented.
Most commentary on AI stops at the diagnosis.
Naming what is wrong with generative AI has become easy, and nearly everyone is doing it. What stays scarce is anyone turning that diagnosis into something a researcher, a teacher or a policy officer can actually use on a Tuesday morning. What follows is the whole argument — what is broken, what I built in response, and the record standing behind both.
What is actually broken
The verification gap, missing provenance, integrity and equity in publishing, the disclosure vacuum, data exposure, and judgement delegated by accident — real-world challenges, each one drawn rather than merely described.
Read the challenges → Eleven systemsWhat I built in response
Screening pipelines, journal integrity tools, a live evidence search portal, and three everyday productivity utilities. Eight are publicly available today, each sitting directly beneath the challenge it answers.
See the systems → The recordWhy I can be useful to you
Evidence reviews, research integrity and publishing governance, AI policy and capability, assessment redesign, and bespoke builds — each grounded in something specific rather than asserted.
What I draw on →