Agentic colonization: who edits Jamaica's view of the world?

A Jamaican reader asks an assistant what happened in the country today. The answer arrives as a short, orderly briefing. Some stories appear; others disappear. A disputed claim may become a confident sentence. The reader saves time, but also hands over part of the editorial work that determines what deserves attention.

That exchange is a useful place to examine agentic colonization, the term Adrian Dunkley uses for dependence on externally controlled AI agents that can transfer practical decision-making authority. It is an analogy about dependency, not an equation with the violence and dispossession of historical colonialism. The question for Jamaica is how much authority people can delegate while keeping their ability to question, refuse and choose.

The risk reaches beyond an inaccurate answer. If the same commercial systems select our news, explain our history, rewrite our language and suggest our next purchase, they can become intermediaries between Jamaicans and their own experience. That possibility deserves scrutiny even when individual tools are useful and their providers have no coordinated intention to manipulate anyone.

The missing story matters as much as the wrong sentence

A summary can contain no obvious falsehood and still give an incomplete picture. Imagine a fictional briefing about a transport dispute. It includes an official statement but omits the commuters waiting for service, the workers explaining their position and the uncertainty over what happens next. The omissions change the reader's understanding without requiring an invented quotation.

For a community organisation, the first defence is a visible source trail. Ask which original reports were consulted, when they were published, whose account is represented and what remains unresolved. Follow the links. A list of plausible-looking publication names is not evidence that the assistant read those pages.

Keep a distinction between a report, an allegation, a forecast and commentary. Those categories often collapse in a smooth paragraph. A teacher preparing a current-affairs discussion can ask students to mark each category separately before considering the assistant's interpretation. The exercise makes editorial choices visible instead of treating the generated answer as a finished account.

Patois should not become an error to be corrected

Language carries relationships, humour and memory. A request to make a community interview clearer should not automatically turn the speaker into somebody else. When an assistant replaces Jamaican expressions with generic international prose, the output may read fluently while losing the speaker's meaning or social position.

This is a plausible pathway to cultural dilution, not a claim that every editing model always erases Jamaican language. Test the actual tool and task. Keep the original recording or transcript, obtain the permissions needed for its use, and ask the speaker or a competent local editor to review changes. If the purpose is translation for a wider audience, say so and retain a route back to the original.

A youth group collecting oral histories should decide which materials may be shared, who may license them and whether any platform may use them for model improvement. Public visibility does not settle those questions. A story being available to read is different from the community agreeing to every possible commercial reuse.

When culture becomes a subscription input

Consider a hypothetical community archive that uploads interviews, lyrics and descriptions of local traditions to a tool. The service helps organise them. Later, the community depends on that service to search and interpret its own collection. Even without establishing that the provider trained a model on those files, the archive faces a practical problem if it cannot export the catalogue, annotations and permissions in a usable form.

Dunkley's documented idea of Extracted Intelligence asks who benefits when community knowledge becomes commercial input. Applied here, the useful question is what the archive retains after the subscription ends. Training rights, access rights, copyright ownership and operational portability are separate issues. A reassuring answer about one should not substitute for answers about the others.

Paying for useful software is not itself exploitation. The warning sign is an arrangement in which the community contributes distinctive value, loses meaningful control and must keep paying to recover ordinary access to that value. A fair arrangement can include payment, attribution, limited licences, community oversight and a workable exit.

Persuasion deserves a stronger boundary

Personalised assistance can explain difficult material in language a reader understands. The same capacity creates questions about influence. A system that knows someone's concerns may frame a recommendation around those concerns. An assistant choosing a news source or suggesting a political interpretation should not be treated as a neutral window simply because it speaks politely.

The strongest claim here is a risk claim. This article does not establish a mass-manipulation campaign against Jamaica, nor does it show that a particular provider is secretly directing Jamaican opinion. Proving such a claim would require evidence of actors, actions, reach and effects. The absence of that proof does not prevent schools and community groups from setting sensible boundaries around profiling and automated persuasion.

Keep sensitive personal histories out of casual prompts. Ask whether an answer contains an advertisement, a paid placement or an affiliate recommendation, and verify any disclosure through the service's own information. For consequential decisions, seek a source with an identifiable human responsibility instead of accepting conversational confidence as accountability.

A seven-day community news audit

The following is an editorial exercise proposed for this article, not a validated research instrument or a set of steps attributed to Dunkley.

1. Choose one local question. Use a matter with accessible public sources, such as a public-service announcement. Avoid entering private case details.
2. Build a human baseline. Before asking the assistant, record 5 facts supported by original sources and 2 questions that remain open. Save the links and dates.
3. Ask for a briefing. Request sources, uncertainty and a separation between reported facts and interpretation. Save the prompt and the complete output.
4. Check 10 statements. Within 7 days, mark each as supported, unsupported, misleadingly framed or unverifiable. Record omissions separately; do not count an omission as a fabricated fact.
5. Change one instruction. Ask for missing perspectives and repeat the same check. Compare the results without claiming that this small exercise measures performance on all Jamaican news.
6. Agree an editorial boundary. Decide which material can be summarised automatically and which requires a named reviewer before it is circulated.

The result is a practical record of where a particular workflow needs supervision. If the second prompt improves the answer, keep the instruction. If the assistant invents sources or repeatedly misses context, narrow its role. It may still be helpful for formatting a verified briefing even when it is unsuitable for selecting the news itself.

Risks and useful limits

The reader bears the immediate cost of a misleading briefing, while the people described may bear reputational harm. A youth speaker can lose their voice through over-editing; an archive can lose bargaining power through poor licensing; a community group can spread an unsupported allegation at speed. Named review, original-source checks, restricted permissions and exportable records address different parts of that problem. None guarantees safety. Human editors also make mistakes, and a small group may struggle to afford review. Start with the material most likely to affect someone's rights, reputation, money or access to services.

AI can help translate, organise and explain. Jamaica benefits when those gains increase people's ability to participate. The practical test is whether the reader remains able to inspect the evidence and disagree with the recommendation.

Questions for community groups

Should we stop using AI summaries?

No blanket rule follows from this argument. Match the task to the consequences. Summarising a verified meeting agenda needs less scrutiny than explaining an allegation about a named person. Keep a human responsible for the final account and give readers the original material where appropriate.

Can a cultural archive publish everything openly?

That is a community decision requiring attention to contributors, ownership, privacy and the intended licence. Some materials may be suitable for open access while others need restricted use. Document the decision for each collection instead of assuming one permission covers every recording or future use.

How can a school run the audit without profiling pupils?

Use public stories and fictional examples. Do not ask pupils to disclose their political views, family finances or personal difficulties to an assistant. Assess the quality of the evidence-checking process, instead of whether the student agrees with the teacher or the model.

What should a correction include?

Identify the statement that changed, provide the source supporting the correction and update the copies you control. If the earlier version was circulated to a group, send the correction through the same channel. Editing the original alone may leave the mistaken account in circulation.