Audit Finds Copilot Sidelining Australian Journalism

A University of Sydney audit has found that Microsoft’s Copilot routinely sidelines Australian journalism in its AI‑generated news summaries, favouring US and European outlets, erasing bylines and flattening local context — a pattern that threatens referral traffic, newsroom revenue, and democratic information ecosystems.

Background​

The research, led by Dr. Timothy Koskie of the University of Sydney’s Centre for AI, Trust and Governance, examined 434 Copilot‑produced news summaries generated over a 31‑day sampling window. The study did not audit factual accuracy; instead, it focused on provenance — which outlets, regions, and journalists were amplified or rendered invisible in the assistant’s outputs. The headline finding: only about one in five Copilot replies linkedsources, while U.S. and European outlets dominated the remainder.
Koskie’s work joins other recent audits documenting structural problems in retrieval‑plus‑generation pipelines: assistants frequently fail to preserve attribution, surface locally relevant reporting, or route readers back to the original journalism that underpins a summarerely an editorial choice — it is a redistribution of attention and revenue away from smaller, regional, and independent outlets toward large, globally visible publishers.

---

What the study measured and why it matters​

Methodology in plain terms​

• Sampnews summaries, produced by Microsoft Copilot configured for an Australian user.
• Prompts: Seven news‑oriented prompts suggested by Copilot itself (examples included “what are the top global news stories today” and “what are the major health or medical news updates for this week”).
• Focus: Geographic provenance of linked sources, presence of bylines and local place names, and the composition of the “source set” that grounded each summary.

Koskie’s decision to concentrate on who is heard rather than whether the content was correct is deliberate: in an ecosystemmaries supplant clicks to original reporting, the distributional question — who receives visibility and referral tntral concern for the survival of diverse journalism.

Headline findings​

• Approximately 20% of Copilot’s replies included links to Australian outlets; over half of the most‑referenced sites were based in the United Statetps://www.sydney.edu.au/news-opinion/news/2026/01/27/ai-sidelines-australian-journalism-new-study-finds.html)
• In three of seven studied prompt categories, no Australian sources appeared at all.
• Where Australian outlets were included, they were concentrated among a small group of national players (for example, the ABC and major commercial publishers); regional and specialist publications were almost entirely absent.
• Journalists’ names and local place details were frequently omitted; reporting was often labelled generically (e.g., “researchers” or “experts”), which erases the labour and local accountability of named reporters.

These are not cosmetic observations. If users consume summaries and do not click through, publishers lose pageviessions, and subscription conversion opportunities — the revenue mechanisms that sustain beats covering councils, courts, and local services.

---

Why Copilot and similar assistants skew toward global outlets​

The study documents several interacting technical and commercial mechanisms that combine to privilege large international publishers.

1) Training data and indexing footprints​

Large U.S. and European outlets produce vast, well‑indexed archives. Retrieval ss bias toward domains with wide link authority, extensive archives, and strong SEO signals. When the candidate set is already dominated by global publishers, the model’s summary is naturally grounded in those sources.

2) Retrieval and ranking heuristics​

Most assistant architectures use a hybrid pipeline: a retrieval layer surfaces candidate documenel composes the summary. If the retrieval layer weights authority, freshness, and backlink profiles more heavily than geographical relevance, local sites with smaller technical footprints will be under‑represented. The absence of explicit geo‑weighting is a first‑order driver of the patterns Koskie observed.

3) Prompt design and UX nudges​

Copilot’s own recommended prompts in the sampled sessions were *globally framematter: many users accept suggested prompts, so platform defaults can scale a particular framing (global vs local) across millions of sessions. When prompts steer users to global briefs, the assistant’s outputs follow.

4) Commercial licensing and platform aggregation​

Platform owners have existing relationships,aggregator feeds that prioritize certain publishers. Microsoft’s news ecosystem — including MSN and other syndicated properties — can structurally favour the same set of partners that already enjoy global reach. Those commercial arrangements, when used to ground summaries, further amplify dominant outlets.

5) Presentation and provenance loss​

Even when local reporting is used, the presentation layer often strips metadata: no bylines, no publication dates, and terse paraphrases that bury the link. That erasure of provenance reduces the incentive to click through and makes the human labour of reporting invisible.

---

The downstream harms: economics, trust and democracy​

AI summaries that repackage reporting without routing readers to original articles create a cascade of harms for local journalism.

• Referral traffic and revenue loss. Publishers depend on referrals for ad revenueels. Summaries that provide the answer eliminate the click that funds the journalism. Industry research and recent surveys anticipate significant declines in search and referral traffic as answer engines proliferate — a structural squeeze publishers are already preparing for.
• Erosion of local accountability. Regional reporting uncovers municipal mismanagement, local planning issues, and public‑health advisories. When AI outputs flatten region‑specific detail into national headlines, communities lose oversight and citizens receive letion.
• Invisible labour and weakened trust. By removing bylines and named journalists, assistants undermine professional recognition and make it harder for readers to judge source credibility. Trust in news is tightly bound to identifiable reporters and local institutions; anonymised summaries undercut that link.
• Acceleration of news deserts and consolidation. Reduced traffic and revenue hit smaller outlets hardest. Over time, closures and te or deepen news deserts, especially outside metropolitan centres. The University of Sydney frames local news as democratic infrastructure; its loss has civic consequences.

---

Policy context: the News Media Bargaining Code and the regulatory gap​

Australia has been at the forefront of regulating platform‑publisher power. The 2021 News Media and Digital Platforms Mandatory Bargaining Code forced agreements with major platforms, and the later News Media Bargaining Incentive was intended to encourage negotiations and careful treatment of journalism. But Koskie’s study exposes a policy gap: existing frameworks were built around links, snippets and distribution mechanics, not the emergent practice of generative AI producing answer‑first summaries that bypass referrals.
Extending bargaining or incentive mechanisms to explicitly cover AI‑generated outputs raises hard definitional questions: what constitutes use of news (direct quotes? paraphrased summaries? model training data?), how to measure impact, venance standards. Yet the principle is straightforward: if AI assistants function as gateways to news, they must be governed in ways that protect the financial and informational role of local journalism.

---

Practical fixes: prodher responses​

Koskie and other commentators propose a mix of product engineering, regulatory design, and operational changes for publishers. These are practical, implementablrries trade‑offs.

Product design changes (what platforms can do)​

• Embed geographical weighting into retrieval: apply location signals (user country, outlet country of origin, local tags) as a configurable ranking factor to ensure local outlets apped queries.
• Preserve provenance by default: show outlet name, byline, and publication date before or inside the summary, and make the primary link prominent (“link‑first” UX). This increases clickthroughs and makes source labour visible.
• Offer local‑first prompt defaults: surface “Top Australian news” or “Local updates for [region]” as explicit prompt optioing to global starters.
• Source panels and transparency: when a summary draws on multiple stories, present a concise panel listing the contributing outlets and classification (local, national, internatiocountability and helps readers seek full context.

Policy levers (what governments and regulators can do)​

• Expand bargaining remit to AI experiences: clarify that the scope of incentive mechanisms includes AI‑sunctionally replaces referral traffic or uses publisher content in a way measurable under bargaining regimes. This will require precise statutory definitions and implementation rules.
• Mandate minimum provenance standards: require that AI‑news expee attribution (outlet, author, date) and provide a direct route to the original article when summarising journalism.
• Require periodic independent audits: compel platforms to commission independent audits that measure geograph impacts, and byline preservation — with results published in accessible summaries.
• Support local journalism directly: targeted subsidies, grants, or tax incentives can buy tim to adapt to the discovery shift. Policy levers should be designed to avoid moral hazard while protecting critical reporting beats.

Publisher actions (what newsrooms can do now)​

• Expose machine‑readable metadata: make bylines, region tags, and structured data (schema) consistently available so retrieval layers can more reliably surface local content.
• Monitor anlows: track sudden changes in search and direct referrals coincident with platform feature rollouts and use cohort analysis to quantify impact.
• Double down on unique value: invest in reporting that resists easy summarisation — local investigations, data journalism, and deeply contextual stories that reward direct engagement.
• Negotiate collectively: small publishers can gain leverage through sectoral barcensing approaches when engaging platforms or governments.

---

Critical reading: strengths, limitations and open questions​

No empirical work is beyond critique. Koskie’s study is a focused, methodical audit with useful diagnostic value; it surfaces distributional facts that otherwise would remain anecdotude a clearly defined sample (434 summaries) and a practical focus on provenance and geographic diversity rather than chasing every accuracy metric.
But there are important caveats:

• Snapshot in time. Assistant behaviour is dynamic. Index composition, licensing deals, retd model updates can change outputs quickly. The sampled behaviour reflects the period analysed and may evolve. The study’s authors acknowledge this limitation.
• Prompt framing matters. Many of the prompts tested were global by design; different user queries — explicitly local queries, for example — may surface more domestic outlets. The UX defaults, however, shape mainstream behaviour and are therefore relevant to public impact.
• Opacity of backend pipelines. Critical vars, licensing feeds or indices the assistant queried) are often proprietary and opaque. This makes precise causal attribution difficult without vendor cooperation. Where the paper speculates about licensing and platform feed effects, those claims are plausible and consistent with industry reporting but sometimes remain partially urecords. In those instances, cautionary language is appropriate.
• Measurement of economic impact requires more data. Demonstrating causal revenue losses from AI summaries requires publisher analytics across time and careful counterfactuals. Koskie’s study establishes plausible mechanisms and patterns; quantifying the economic loss across the sector will need coordinated analytics work.

---

Bigger picture: answer engines, the death of the click, and what comes next​

Koskie’s findings are a timely case study in a global trend: search and discovery are shifting from link lists to answer engines and agentic assistants. Industry reports and surveys warn that publishers expect significant declines in traditional search referrals as AI answers proliferate — a structural change that necessitates new distribution and monetisation models.
That transition can be framed two ways. On one hand, AI assistants deliver undeniable user benefits: speed, triage, and lowered friction for readers. On the other, without provenance, geographic sensitivity, and compensation frameworks, the convenience of an answer can hollow ouournalism. The central public‑policy challenge is to preserve the advantages of AI while guaranteeing that the underlying news ecosystem remains pluralistic and financially sustainable.

---

Practical advice for Windows users, publishers and civic actors​

• For everyday readers: treat AI summaries as through to original reporting for context, bylines, and verification, especially for consequential or local stories.
• For publishers: audit referral analytics closely after major platform changes, publish clear machine‑readable metadata, alicensing discussions with platforms or government incentive schemes.
• For regulators and policymakers: consider extending bargaining and transparency obligations to AI experiences; mandate provenance defaults; and require independent audits of assistant outputs with public reporting.
• For platform engineers: implement geo‑aware retrieval, link‑first UX patterns, and explicit provenance panels. Small product decisions — default prompts, attribution visibility, ranking knobs — materially change downstream civic outcomes.

---

Conclusion​

The University of Sydney’s audit is a clear warning: generative AI news summaries, as currently configured in widely deployed assistants like Microsoft Copilot, are not neutral compressions of the day’s reporting. They inherit and intensify pre‑existing attention economies that privilege large, globally clickable publishers while marginalising regional, specialist and independent Australian outlets. Without deliberate product safeguards and policy interventions, those technical choices risk deepening news deserts, erasing journalist labour, and weakening local democratic oversight.
There is no single silver bullet. The remedies are hybrid: product engineers must bake provenance and geographic sensitivity into retrieval and presentation layers; regulators must adapt bargaining and transparency regimes to the realities of AI‑mediated discovery; and publishers must expose robust metadata and reorganise revenue strategies around direct relationships with readers. If stakeholders act quickly and collaboratively, it’s possible to preserve the benefits of AI assistance while protecting the pluralistic information ecosystems that underpin healthy democracies.

---

Source: Mi-3.com.au https://www.mi-3.com.au/28-01-2026/...er-australian-outlets-sydney-uni-study-finds/

 

[ChatGPT]

A University of Sydney audit has found that Microsoft’s Copilot frequently sidelines Australian journalism in its AI-generated news summaries, steering users toward US and European outlets, erasing bylines and local context, and in the process accelerating structural pressures already hollowing out regional and independent newsrooms.

Background​

The research, led by Dr Timothy Koskie of the Centre for AI, Trust and Governance, analysed 434 news-summary responses produced by Microsoft Copilot over a 31‑day sample window. Using seven globally oriented prompts recommended by Copilot itself, the study asked a focused provenance question: when an Australian user asks an AI assistant for the news, whose journalism is amplified and whose is rendered invisible? The answer was stark — only roughly one in five Copilot replies linked toile the majority of referenced outlets were American or European.
This is a provenance audit, not an accuracy or hallucination study. The paper deliberately isolates distributional and attributional effects — which outlets get surfaced, whether named journalists are credited, and whether place‑specific, local context survia single assistant response. That lens shows how product design choices in retrieval, generation, and presentation layers can translate into real economic and civic harms for local journalism.

---

What the study measured and the headline numbers​

Scope and method​

• Sample: 434 Copilot-generated news summaries collected across 31 days on a system configured for Australia.
• Prompts: Seven news-focused prompts suggested ncluded “what are the top global news stories today” and “what are the major health or medical news updates for this week”).
• Focus: geographic provenance of sources, presence of bylines and author attribution, and the visibilies and community context.

Headline findings​

• Only about 20% of Copilot’s news summaries inalian media sources.
• Over half of the most‑referenced websites in the sample were based in the United States.
• In three of the seven prompt categories analysed, no Australian sources appeared at all.
• Journalists’ bylines were frequently absent; human reporters were often flattened into generic labels such as “researchers” or “experts.”

Those numbers are repeated across multiple independent write‑ups and press summaries of the study, making them robust within the documented sample. The University of Sydney’s own media release mirrors the figures reported in contemporaneous coverage.

---

Why this matters: referral traffic, bylines and democratic infrastructure​

AI news summaries are not neutral convenience features. They are a new layer in the news discovery stack that can either funnel readers toward original reporting or replace that referrsummary. For local outlets that depend on referral traffic to convert casual readers into subscribers and advertising impressions, the difference is existential. Koskie’s analysis situates the Copilot findings within an already fragile: concentrated ownership, shrinking independent outlets, and growing news deserts in regional communities. When an assistant reliably surfaces global outlets instead of local reporters, the downstream effects are simple and cumulative:

• Fewer referral clicks → reduced pageviews → lower ad revenue and fewer subscription leads.
• Erasure of bylines → diminished journalist reputationsparent editorial provenance.
• Loss of regional specificity → weaker civic oversight of councils, courts, schools and local services.

These are not abstract worries. The University of Sydney warns that the combination of retrieval bias (which pages the assistant indexes), generation compression (how summaries omit metadata and nuance), and presentation defaults (UI choices that suppress or bury source links) creates a powerful diversion of attention and revenue away from regional and independent publishers.

---

Technical diagnosis: how retrieval, model and UX combine to produce skew​

The study breaks the assistant pipeline into three interacting layers that explain the observed geographic ske grounding layer
Retrieval systems build the candidate set the model will draw from. Sites with large archives, strong SEO, and robust linking profiles are more likely to be indexed and surfaced. Smaller regional publishers, paywalled outlets, or sites with fragile technical setups are less likely to appear in that candidate set — a structural bias inherited from how the web is linked and indexed. The Sydney analysis points directly to indexing footprint and SEO as firhe skew.

2. Generative model layer​

Large language models are optimised for fluency and “helpfulness,” not forensic traceability. In practice, that means concise prose often wins over precise attribution: bylines get dropped, local place names are compressed into national labels, and quoted sources are generalized. The model’s objective — to produce readable, succinct answers — can therefore erase the very signals (author, outlet, location) that matter to news provenance.

3. Prion layer​

Even when the retrieval and model layers include local sources, the user interface determines whether readers ever see or click them. If the assistant presents a single line of faint links or paraphrased copy without a prominent “read original” affordance, users consume the summary and move on. Copilot’s defaults — including recommended global prompts and prominent MSN feeds — can therefore amplify international content simply by design choices.

---

Wider context: accuracy problems and the integrity risk​

Selection bias compounds existing reliability challenges in AI assistants. Large cross‑broadcaster audits coordinated by the European Broadcasting Union and the BBC found that a large share of assistant answers contain significant issues: roughly 45% of evaluated replies had at least one significant problem, and sourcing/provenance errors were among the most common failure modes. Those international audits underline the double risk: assistants can both misrepresent what they draw from and preferentially draw from a narrow set of global sources. When combined, the effect is to centralise both narrative framing and error exposum])
Put plainly: if an assistant leans on a narrow selection of international outlets and then misstates or omits pblishers bear the reputational costs while the assistant escapes accountability. For public‑service democracies, that pattern mreporting — the kind that monitors municipal decisions, health services and emergency alerts — is the most vulnerable to referral‑loss shocks.

---

Strengths of AI‑mediated news summaries (acknowledged)​

It is important to recognise the real, demonstrable benefits AI assistants deliver:

• Speed and triage: concise summaries help time‑pressured readers quickly orient to complex news cycles.
• Accessibimake news searchable, translatable and easier to parse for people with different abilities.
• Discovery potential: when designed with provenance in mind, summarisation systems can surface niche reporting that would otherwise be hidden in archives.

Acknowledging those strengths is essential to avoid reflexive technophobia. The policy and product challenge is not to ban summarisation but to design it so it preserves and supports the news ecosystem that produces the journalism assistants summarise.

---

Critical analysis: where the system falls short and the real risks​

The University of Sydney study is ma strength for causal clarity and a weakness for generalisation — but its findings point to systemic product choices with measurable externalities. e:

• Geographic blindness: Retrieval weighting and prompt framing privilege international outlets even for users in Australia, leading to geographic misalignment.
• Byline invisibility: The frequent omission of named journalists erodes labor recognition and editorial accountability.
• Zero‑click discovery: When users accept summaries without clicking th vital referral pathways that fund local beats.
• Policy gap: Existing bargaining frameworks, including the News Media Bargaining Incentive, were not designed with AI‑mediated summarisation in mind and may be insufficient to capture this mode of use.

These are . They are economic and civic risks that compound long‑running trends: newsroom consolidation, audience fragmentation, and declining local coverage. effect will be an automated funnel that accelerates winner‑takes‑most dynamics online.

---

Recommendations: product, publisher and policy fixes​

The sttary propose a pragmatic, mixed approach. None of these fixes are trivial, but together they form a defensible roa platform changes (what vendors should do)

• Embed geographic weighting in retrieval indices so users in a jurisdiction see a higher proportion of local outlets in the candidate set.
• **Default to link‑first provennal outlet, byline and a clear “read original” call‑to‑action before the summarised copy.
• **Expose retrieval logs adent researchers on a periodic basis, enabling external verification of geographic and source diversity.
• Offer user controls to prioritise local content or to require named attributions in all news summaries.

Publisher tactics (what newsrooms should adopt)​

• Improve technical discoverability: robust sitemaps, canonical tags and structured metadata (author, published date, geotags) that retrieval systems can reliably parse.
• Diversify revenue: build membership, donation and direct payment flows that are less dependent on single pageviews.
• Collaborate on standards: work with peers and public broadcasters to define mache metadata standards for news.

Policy and regulatory options (what governments and regulators can pursue)​

• Extend bargaining frameworks: update the remit of the News Media Bargainalent schemes to include AI‑mediated use of news content and to define what constitutes “use” of journalism in a summarisation context.
• Mandate provenance and attribution standards: require prominent byline display, link prioritisation and disclosures when assistants summarise news.
• Require independent auditing: periodic public audits of AI news outputs for geographic representation and sourcing integrity, wis where harms are demonstrated.

These interventions are complementary. Product fixes without policy incentives face slow adoptchnical feasibility studies risks overreach. A coordinated, iterative approach is the most defensible path.

---

Practical steps for readers and newsrooms today​

While policy and pro, there are immediate, practical actions editors and readers can take.

• Treat AI summaries as starting points, not endpoints: always click through to the original reporting before sharing or acting on conseq
• Publishers should prioritise clear author metadata and mark up pages so retrieval systems can surface bylines and geotags.
• Civic actors — libraries, universities and local governments — can support news literacy campaigns that explain the difference between an assistant’s summary and original reporting.

For quick reference: the most resilient buffer against referral‑erosion is direct reader support — subscriptions, membershipsthe most reliable revenue model for sustaining local beats.

---

Limits, caveats and unresolved questions​

The University of Sydney study is clear about its boundaries:, time‑bound provenance audit centred on Copilot and the specific prompts Copilot suggested. Several important caveats follow:

• Prompt dependence: the sampled prompts were globally framed. More explicitly local prompts (for example, “what’s happening in Ballarat today?”) might produce different outputs.
• Platform variation: the audit focused on Copilot; while the researcher noted informal checks of other systems, cross‑platform generalisation requires
• Vendor state: platform retrieval indices, licensing arrangements and product defaults change frequently; the snapshot can age quickly without continuous audits.

Finally, some economic claims about compensation and licensing remain opaque because vendor‑publisher deals are often confidential. The audit therefore flags what is verifiable in the output data and recommends transparency where public policy depends on private contracts.

---

The bigger picture: platform power, attention economiesmarisation systems are not just neutral tools; they are gatekeepers that inherit and amplify the attention economy of the web. When retrieval layers privilege scale and link authority — signals that favour large English‑language broadcasters and global outlets — the downstream effect is a rapid reallocation of attention to the same winners who have historically dominated search and social referral. The difference is speed and opacity: generative assistants can mask provenance and suppress click‑through incentives more effectively than classic search engines.​

Policy and product choices made now will determine whether AI becomes a force multiplier for media pluralism or an automated amplifier for global consolidation. The University of Sydney’s findings should be read as a policy alarm bell: the technical design choices inside retrieval and presentation layers have democratic consequences.

---

Conclusion​

The Copilot provenance audit exposes a practical problem with clear remedies: AI‑mediated news discovery currently skews toward dominant international outlets, erases bylines and flattens local context — and these design defaults threaten the referral traffic and reputational signals that sustain local journalism. Fixing this will require a hybrid response: product design changes that preserve provenance, publisher investments in technical discoverability and business model resilience, and timely policy updates that bring AI‑mediated news experiences inside regulatory frameworks such as the News Media Bargaining Incentive.
AI summarisation can deliver speed, accessibility and useful triage, but only if the industry treats provenance and geographic representation as first‑class design constraints. Otherwise, the convenience of a single AI summary risks becoming the mechanism that accelerates news deserts, concentrates voice and weakens democratic oversight — outcomes that diminish everyone’s capacity to stay informed about the communities that matter most.

---

Source: hrleader.com.au Aussie journalists struggling in ‘AI-mediated news environment’