Mihar Dias’s August 5 column, “The Intern in the Newsroom,” lands on the right operational rule for journalists using generative AI: treat it as a fast, capable assistant whose work is unverified until a human proves otherwise. The problem is that the column’s most important warning is not about fabricated quotations or an imitation of a columnist’s voice. It is about whether newsroom management will spend AI’s promised productivity gains on better reporting—or simply demand more copy from the same number of people.
Published through Newswav and framed in its metadata as a piece on Microsoft Copilot, the column recounts a keynote by Dr. Rahim Said titled “AI for Journalists and Columnists — Writing Smarter Without Losing Your Voice,” reportedly organised by Weekly Echo. Dias argues that AI can draft, summarize, translate and sharpen headlines, but cannot supply the reporting memory, contextual judgement or personal accountability that gives a journalist’s byline meaning.
That is sound advice, but the published piece provides no date, venue, recording, agenda, speaker biography or independently verifiable account of the event. Searches for the exact keynote title and the named event did not surface corroborating coverage. That does not establish that the talk did not occur; it means readers should treat the keynote details as Dias’s account, not as independently confirmed reporting.
The underlying argument, however, is already being tested in newsrooms. The Associated Press updated its AI newsroom standards on July 23, 2026, allowing tightly bounded uses including early research, document summarization, transcription, translation, headline suggestions and grammar support. AP’s condition is blunt: AI output must be reviewed and edited by journalists, and AI does not replace reporting, sourcing, editorial judgment or verification.
Dias’s “brightest intern” metaphor is therefore useful—but incomplete. A good newsroom intern has an identity, can be questioned about a source, may retain institutional learning after correction and can be held accountable for misconduct. A model can produce a plausible answer without knowing whether its supporting claim exists. The right comparison is an intern who can write at industrial speed, has no memory of prior corrections, and may be connected to corporate systems holding material the reporter was never supposed to see.
Dias correctly identifies the pressure point: AI may create time, but business incentives decide where that time goes. If editors use an hour saved on transcription for another interview, checking documents against originals, or calling a skeptical second source, AI can improve a story. If the same hour becomes a higher daily output target, it can multiply unverified reporting while preserving the appearance of human authorship.
That is the practical difference between augmentation and automation. It cannot be settled by telling individual reporters to “use AI responsibly.” A reporter facing a same-day quota, limited editing and a requirement to publish first has little power to convert saved time into deeper reporting. The policy must be owned by the editor who assigns work, the product team that deploys tools and the manager who measures performance.
Malaysian media voices have been making the same point in public. In May, Universiti Malaya lecturer Muhammad Zaiamri Zainal Abidin told Bernama Radio that speed cannot be pursued at the expense of ethics, while calling for clearer AI guidance for the editorial sector. In June, Broadcasting Director-General Ashwad Ismail argued that journalists need to learn to use AI while preserving reporting’s human role. Those statements support adoption, but neither substitutes for a published, enforceable newsroom policy.
A credible policy answers questions that a keynote and a prompt guide cannot:
That omission matters because “Copilot” is not one uniform data-handling environment. A reporter using a consumer chatbot in a browser, an employee using Microsoft 365 Copilot Chat under enterprise data protection, and a Microsoft 365 Copilot user drawing on documents and mail through Microsoft Graph are operating under materially different controls.
Microsoft’s current documentation says that, under enterprise data protection, prompts and responses in Microsoft 365 Copilot and Microsoft 365 Copilot Chat are not used to train foundation models. Microsoft also says those systems apply an organization’s existing permissions, sensitivity labels, retention settings and auditing controls. For an IT administrator, that is a significant advantage over staff freely pasting material into unmanaged tools.
But “not used to train the model” is not the same as “invisible.” Microsoft states that enterprise-protected prompts and responses can be logged, retained and made available for audit and eDiscovery, depending on the subscription and configuration. A journalist or editor working with confidential material therefore needs to understand both sides of that protection: the content is governed inside the organizational environment, and the organization may have a durable record of the interaction.
Microsoft 365 Copilot also respects the permissions model already in the tenant. That limits a user to material they are authorized to access, but it exposes another newsroom problem: AI can reveal existing oversharing faster than a conventional search workflow. If broad SharePoint folders, Teams channels or newsroom archives have loose permissions, Copilot’s ability to summarize and connect material can make those access-control failures more visible and more consequential.
The first Copilot project for a newsroom should therefore be a permissions and retention review, not a headline-generation pilot. Admins should map where unpublished drafts, source-identifying files, legal correspondence, HR documents and embargoed materials live; apply appropriate sensitivity labels; review membership of shared workspaces; and decide what interaction logging means for source protection and legal discovery. The tool cannot fix a tenant whose information governance was already weak.
The Associated Press has described AI output as unvetted source material. Its March guidance on AI and data journalism goes further, stressing transparency, reproducibility and accuracy—areas where generative systems can be weak. AP’s practical questions are the ones an editor should require before accepting AI-assisted analysis: Was AI used? Why was it chosen? How was it applied? How was the result verified?
Those questions are especially relevant to the tasks Dias recommends. Summarizing a public report can be useful, but the journalist must still read the relevant pages and verify every number and quotation used. Translation can accelerate access, but a fluent output should be checked by a competent speaker where wording has legal, political or cultural significance. Headline suggestions may save minutes, but they can still introduce defamation risk or frame a developing story more strongly than the reporting supports.
The same distinction applies to columnists. AI can propose a structure, shorten a sentence or list counterarguments. It cannot supply the experience of reporting a minister’s evasive answer, weighing a source’s credibility after years of contact, or deciding which personal observation is fair to publish. Dias is right that voice is not merely cadence. It is the accumulated record of judgement behind the words.
Still, voice is not the only stake. A polished AI-assisted column can sound distinctly human while containing a false factual premise. Readers cannot audit the drafting process from the finished prose. That places more weight, not less, on the writer and editor to maintain notes, source records and a clear distinction between observed fact, reported fact and opinion.
Measure it instead by whether reporters spend more time in the places models cannot go: obtaining documents, meeting sources, noticing contradictions, challenging official accounts and checking what an apparently authoritative answer leaves out. Those are not nostalgic rituals. They are the controls that distinguish reporting from content generation.
Dias’s column offers a memorable instruction—use the intern, but keep track of what it cannot do. For editors and Microsoft 365 administrators, the operational version is sharper: approve the task, protect the data, preserve the audit trail, verify the output, and refuse to convert saved minutes into an excuse to remove human scrutiny.
If Copilot or another assistant gives a reporter twenty minutes back at the end of the day, the correct destination for those minutes is not another generated article. It is the phone call, document check or second interview that makes the published one worth trusting.
That is sound advice, but the published piece provides no date, venue, recording, agenda, speaker biography or independently verifiable account of the event. Searches for the exact keynote title and the named event did not surface corroborating coverage. That does not establish that the talk did not occur; it means readers should treat the keynote details as Dias’s account, not as independently confirmed reporting.
The underlying argument, however, is already being tested in newsrooms. The Associated Press updated its AI newsroom standards on July 23, 2026, allowing tightly bounded uses including early research, document summarization, transcription, translation, headline suggestions and grammar support. AP’s condition is blunt: AI output must be reviewed and edited by journalists, and AI does not replace reporting, sourcing, editorial judgment or verification.
Dias’s “brightest intern” metaphor is therefore useful—but incomplete. A good newsroom intern has an identity, can be questioned about a source, may retain institutional learning after correction and can be held accountable for misconduct. A model can produce a plausible answer without knowing whether its supporting claim exists. The right comparison is an intern who can write at industrial speed, has no memory of prior corrections, and may be connected to corporate systems holding material the reporter was never supposed to see.
The Missing Policy Is More Important Than the Prompt
Dias correctly identifies the pressure point: AI may create time, but business incentives decide where that time goes. If editors use an hour saved on transcription for another interview, checking documents against originals, or calling a skeptical second source, AI can improve a story. If the same hour becomes a higher daily output target, it can multiply unverified reporting while preserving the appearance of human authorship.That is the practical difference between augmentation and automation. It cannot be settled by telling individual reporters to “use AI responsibly.” A reporter facing a same-day quota, limited editing and a requirement to publish first has little power to convert saved time into deeper reporting. The policy must be owned by the editor who assigns work, the product team that deploys tools and the manager who measures performance.
Malaysian media voices have been making the same point in public. In May, Universiti Malaya lecturer Muhammad Zaiamri Zainal Abidin told Bernama Radio that speed cannot be pursued at the expense of ethics, while calling for clearer AI guidance for the editorial sector. In June, Broadcasting Director-General Ashwad Ismail argued that journalists need to learn to use AI while preserving reporting’s human role. Those statements support adoption, but neither substitutes for a published, enforceable newsroom policy.
A credible policy answers questions that a keynote and a prompt guide cannot:
- AI-assisted research must never be treated as a source; every factual assertion used in copy must be traced to a primary document, a named source, or independently reliable reporting.
- Reporters must not paste confidential source material, unpublished allegations, legal drafts, victim information, embargoed documents or private interview transcripts into an unapproved public AI service.
- An editor must know when generative AI materially shaped a published piece, particularly where it created substantive prose, analysis, images, audio or translations.
- Performance targets must not assume that automated drafting eliminates the reporting and editing time needed to validate output.
“Copilot” Is Named, but the Product Details Are Absent
There is another gap worth noting for Windows and Microsoft 365 users. The Newswav presentation calls the piece “Mihar Dias on Microsoft Copilot,” but the column itself does not identify a Copilot edition, a Microsoft 365 license, a tenant configuration, or a workflow using Microsoft software. It discusses “AI” and “language models” generally.That omission matters because “Copilot” is not one uniform data-handling environment. A reporter using a consumer chatbot in a browser, an employee using Microsoft 365 Copilot Chat under enterprise data protection, and a Microsoft 365 Copilot user drawing on documents and mail through Microsoft Graph are operating under materially different controls.
Microsoft’s current documentation says that, under enterprise data protection, prompts and responses in Microsoft 365 Copilot and Microsoft 365 Copilot Chat are not used to train foundation models. Microsoft also says those systems apply an organization’s existing permissions, sensitivity labels, retention settings and auditing controls. For an IT administrator, that is a significant advantage over staff freely pasting material into unmanaged tools.
But “not used to train the model” is not the same as “invisible.” Microsoft states that enterprise-protected prompts and responses can be logged, retained and made available for audit and eDiscovery, depending on the subscription and configuration. A journalist or editor working with confidential material therefore needs to understand both sides of that protection: the content is governed inside the organizational environment, and the organization may have a durable record of the interaction.
Microsoft 365 Copilot also respects the permissions model already in the tenant. That limits a user to material they are authorized to access, but it exposes another newsroom problem: AI can reveal existing oversharing faster than a conventional search workflow. If broad SharePoint folders, Teams channels or newsroom archives have loose permissions, Copilot’s ability to summarize and connect material can make those access-control failures more visible and more consequential.
The first Copilot project for a newsroom should therefore be a permissions and retention review, not a headline-generation pilot. Admins should map where unpublished drafts, source-identifying files, legal correspondence, HR documents and embargoed materials live; apply appropriate sensitivity labels; review membership of shared workspaces; and decide what interaction logging means for source protection and legal discovery. The tool cannot fix a tenant whose information governance was already weak.
Verification Is Work, Not a Checkbox
The phrase “hallucination” is often used too loosely in AI discussions. For a newsroom, the concern is more concrete: a generated answer may invent a quotation, reverse a date, combine details from unrelated people, misread a table, omit a qualifying condition, or produce a citation that looks credible but does not support the claim. Each failure mode requires a different check.The Associated Press has described AI output as unvetted source material. Its March guidance on AI and data journalism goes further, stressing transparency, reproducibility and accuracy—areas where generative systems can be weak. AP’s practical questions are the ones an editor should require before accepting AI-assisted analysis: Was AI used? Why was it chosen? How was it applied? How was the result verified?
Those questions are especially relevant to the tasks Dias recommends. Summarizing a public report can be useful, but the journalist must still read the relevant pages and verify every number and quotation used. Translation can accelerate access, but a fluent output should be checked by a competent speaker where wording has legal, political or cultural significance. Headline suggestions may save minutes, but they can still introduce defamation risk or frame a developing story more strongly than the reporting supports.
The same distinction applies to columnists. AI can propose a structure, shorten a sentence or list counterarguments. It cannot supply the experience of reporting a minister’s evasive answer, weighing a source’s credibility after years of contact, or deciding which personal observation is fair to publish. Dias is right that voice is not merely cadence. It is the accumulated record of judgement behind the words.
Still, voice is not the only stake. A polished AI-assisted column can sound distinctly human while containing a false factual premise. Readers cannot audit the drafting process from the finished prose. That places more weight, not less, on the writer and editor to maintain notes, source records and a clear distinction between observed fact, reported fact and opinion.
The Better Measure of AI’s Value
The value of AI in journalism should not be measured by how many stories it helps a newsroom produce. That metric rewards the least defensible use of the technology: publishing more quickly than the organization can verify.Measure it instead by whether reporters spend more time in the places models cannot go: obtaining documents, meeting sources, noticing contradictions, challenging official accounts and checking what an apparently authoritative answer leaves out. Those are not nostalgic rituals. They are the controls that distinguish reporting from content generation.
Dias’s column offers a memorable instruction—use the intern, but keep track of what it cannot do. For editors and Microsoft 365 administrators, the operational version is sharper: approve the task, protect the data, preserve the audit trail, verify the output, and refuse to convert saved minutes into an excuse to remove human scrutiny.
If Copilot or another assistant gives a reporter twenty minutes back at the end of the day, the correct destination for those minutes is not another generated article. It is the phone call, document check or second interview that makes the published one worth trusting.
References
- Primary source: Newswav
Published: 2026-08-05T03:00:50+00:00
Loading…
newswav.com - Related coverage: newswav.com
Loading…
newswav.com - Related coverage: ap.org
Loading…
www.ap.org - Related coverage: journalists.org
Loading…
www.journalists.org - Related coverage: awards.journalists.org
Loading…
awards.journalists.org - Related coverage: ona26live.journalists.org
Loading…
ona26live.journalists.org - Related coverage: journalists.org
Loading…
www.journalists.org - Related coverage: ona19.journalists.org
Loading…
ona19.journalists.org - Related coverage: learn.microsoft.com
Microsoft 365 Copilot Chat Privacy and Protections | Microsoft Learn
Microsoft 365 Copilot Chat protects workplace AI-powered web chats by providing enterprise data protection to keep organizations safe. Learn about the data protections, authentication, authorization, and GDPR compliance.learn.microsoft.com - Related coverage: learn.microsoft.com
Microsoft 365 Copilot Chat Privacy and Protections | Microsoft Learn
Microsoft 365 Copilot Chat protects workplace AI-powered web chats by providing enterprise data protection to keep organizations safe. Learn about the data protections, authentication, authorization, and GDPR compliance.learn.microsoft.com - Related coverage: support.microsoft.com
Frequently asked questions about Microsoft 365 Copilot Chat | Microsoft Support
Get answers to common questions about what Microsoft 365 Copilot Chat, your AI assistant for work, can do.support.microsoft.com - Related coverage: techcommunity.microsoft.com
Loading…
techcommunity.microsoft.com