The insight is what matters

A mistake can become a comforting story

A chatbot gives a ridiculous answer. Someone shares the screenshot. The comments turn it into a verdict on artificial intelligence, and for a moment the future feels less threatening.

I understand the relief. When a tool starts doing work that took us years to learn, it touches our sense of identity. Finding its mistakes can feel like recovering some ground.

Those mistakes deserve attention. They help us identify where a system needs better evidence, supervision or a different method. I lose interest when the mistake becomes a reason to dismiss everything the tool can do.

My practical starting point is to use AI where it improves the work. That may mean searching documents, comparing alternatives, drafting code or organising an argument. The useful question is how well a particular system performs a particular task, and what it takes to check the result.

Research offers a useful example. In a study involving 758 consultants, AI assistance improved performance on tasks within the system’s capabilities, while performance suffered on a task outside them. The authors describe an uneven boundary of capability, a boundary that has to be understood through the work itself. Dell’Acqua and colleagues, Harvard Business School.

That is the boundary I want to keep exploring as the tools change.

The human role moves with the tools

I see useful human work wherever there is a decision about purpose, a missing observation, a physical constraint or a person whose needs are poorly represented by the available data.

Some tasks remain difficult because the technology cannot handle them reliably. Others require permission, trust or a named person who can answer for the outcome. These conditions can change at different speeds.

In robotics, that distinction becomes tangible. A model may propose a motion. An engineer still has to decide how to build the mechanism, gather the right measurements and establish whether the movement is suitable around people. Every connection between a model, a machine and a person creates another place for judgement.

I want to work at those connections. As AI takes on more of the analysis, human attention can move towards the conditions that make an action useful, acceptable and worth taking.

There is also a limit to defining our place through productivity. A person has a life to experience, relationships to sustain and reasons to create. Those matter even when a machine can produce an impressive result.

What an AI detector can tell me

When I read a technical article, I want to know what the author noticed. I want to understand the evidence, follow the reasoning and see whether the conclusion survives a serious question.

AI can help formulate those insights. It can also suggest something the author had not considered. The person putting their name on the text still chooses what to endorse, what to verify and what to do with it.

A detector label gives me a different kind of information: an estimate about the origin of the language. Its usefulness depends on the question being asked. An assessment of independent writing, a disclosure requirement and a search for a useful engineering idea each call for different evidence.

For judging the insight itself, I find a detector score a poor substitute for reading the argument.

One passage, a structural rewrite and a human label

I explored this through a technical passage from research on soft robotic capture. An AI assistant produced two revisions. The rewrite used in the second attempt reorganised the passage into clearer paragraphs.

The revision separated the timing problem, the capture criterion and the simulation results. It used more direct subjects and verbs. Here is one short comparison between the first attempt and the second attempt.

First attempt, original passage

Success is defined as safe trapping, in which the ball inscribed in the satellite cannot leave the arms, whatever the final pose. Evaluated instant by instant, the same criterion defines a caging time.

Second attempt, structural rewrite

We assess capture through safe trapping. Whatever the final pose, the ball inscribed in the satellite must be unable to leave the arms. Evaluating this condition instant by instant defines a caging time and yields design conditions normalized by the target side.

The extra design-condition clause in this excerpt from the second attempt comes from the next sentence of the original. That is one example of how the assistant regrouped the reasoning. The full texts contain 439 words for the first attempt and 427 for the second attempt. The experiment’s checks recorded preserved numeric tokens and protected technical terms; scientific equivalence still requires a reading of the claims and their conditions.

The result supplied for the second attempt shows Pangram 4.0 classifying 100% of the analysed text as Human Written, with 438 words reported as scanned. The rewrite was composed by the assistant before that result was known.

Recorded observation	Displayed result	Context
First attempt	46% AI	449 words reported
Second attempt	100% Human Written	438 words reported

The original from the first attempt was not rescanned alongside the second attempt. The different word counts also leave a question about extraction or counting by the service. Scan identifiers and complete service exports are missing. These observations do not support a measured improvement of 46 percentage points or a general bypass rate.

They do document a useful case: a passage rewritten by AI received a fully human label. Knowing how the revision was produced gives me stronger evidence of its AI involvement than that label does.

The folder holding the experiment is named GPTZero. The service shown in this result is Pangram 4.0.

The label and the history of the text

The file list also shows three labels, AI for a version with a neutral preface, Mixed for a version containing a quoted instruction, and Human for the rewrite used in the second attempt. The prefaces and prose differ between those files, so the list illustrates variation across inputs rather than a controlled three-stage improvement.

Separate paired tests compared commands with quoted instructions. They produced 37% versus 34% AI in one pair and 32% versus 30% in the other. The predicted pattern of following the requested classification did not appear. Those results belong in the record too.

My strongest conclusion is about the weight given to the output. A label cannot establish a text’s history on its own. Earlier research has also shown that paraphrasing can reduce detection in the systems tested. Those studies provide context, while their results remain specific to their models and test conditions. Sadasivan and colleagues, Transactions on Machine Learning Research.

There are people on the receiving end of these judgements. A 2023 study found that several detectors disproportionately flagged writing by people using English as an additional language. It is a reason to examine the evidence behind an accusation, with attention to the detector and population involved. Liang and colleagues, Patterns.

Pangram’s own documentation describes a humanizer detection feature, advises removing extraneous instructions before analysis and recommends text or DOCX over PDF because extraction can introduce artefacts. These details matter when interpreting this small experiment. Pangram 4 model card.

For work that needs an account of how it was produced, I would look at the drafts, the sources, the decisions and the author’s ability to explain them. A detector may contribute a signal within that process. I would give the larger judgement to the evidence.

Leave room for the person

Poetry makes my preference particularly clear. When someone writes about grief, I care about the person who has lived with it. I care about what they choose to reveal, and about the recognition another person may find in those words.

AI can generate moving language. A writer may use it while expressing something deeply personal. The experience they bring and their decision to share it remain part of why I want to read them.

For now, I feel little need to ask a model what it claims to feel. I am more interested in what a person feels, what they notice and what they decide to do. Creativity can come from curiosity, joy, attachment and pain. I want room for people to give those experiences a form of their own.