Freedom of expression

AI still struggles to mimic natural human conversation

A recent study reveals that large language models such as ChatGPT-4, Claude, Vicuna, and Wayfarer still struggle to replicate natural human conversation. Researchers found AI over-imitates, misuses filler words, and struggles with natural openings and closings, revealing its artificial nature.

The research, led by Eric Mayor with contributions from Lucas Bietti and Adrian Bangerter, compared transcripts of human phone conversations with AI-generated ones. AI can speak correctly, but subtle social cues like timing, phrasing, and discourse markers remain hard to mimic.

Misplaced words such as ‘so’ or ‘well’ and awkward conversation transitions make AI dialogue recognisably non-human. Openings and endings also pose a challenge. Humans naturally engage in small talk or closing phrases such as ‘see you soon’ or ‘alright, then,’ which AI systems often fail to reproduce convincingly.

These gaps in social nuance, researchers argue, prevent large language models from consistently fooling people in conversation tests.

Despite rapid progress, experts caution that AI may never fully capture all elements of human interaction, such as empathy and social timing. Advances may narrow the gap, but key differences will likely remain, keeping AI speech subtly distinguishable from real human dialogue.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Public consultation flaws risk undermining Digital Fairness Act debate

As the European Commission’s public consultation on the Digital Fairness Act enters its final phase, growing criticism points to flaws in how citizen feedback is collected.

Critics say the survey’s structure favours those who support additional regulation while restricting opportunities for dissenting voices to explain their reasoning. The issue raises concerns over how such results may influence the forthcoming impact assessment.

The Call for Evidence and Public Consultation, hosted on the Have Your Say portal, allows only supporters of the Commission’s initiative to provide detailed responses. Those who oppose new regulation are reportedly limited to choosing a single option with no open field for justification.

Such an approach risks producing a partial view of European opinion rather than a balanced reflection of stakeholders’ perspectives.

Experts argue that this design contradicts the EU’s Better Regulation principles, which emphasise inclusivity and objectivity.

They urge the Commission to raise its methodological standards, ensuring surveys are neutral, questions are not loaded, and all respondents can present argument-based reasoning. Without these safeguards, consultations may become instruments of validation instead of genuine democratic participation.

Advocates for reform believe the Commission’s influence could set a positive precedent for the entire policy ecosystem. By promoting fairer consultation practices, the EU could encourage both public and private bodies to engage more transparently with Europe’s diverse digital community.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Wikipedia faces traffic decline as AI and social video reshape online search

Wikipedia’s human traffic has fallen by 8% over the past year, a decline the Wikimedia Foundation attributes to changing information habits driven by AI and social media.

The foundation’s Marshall Miller explained that updates to Wikipedia’s bot detection system revealed much of the earlier traffic surge came from undetected bots, revealing a sharper drop in genuine visits.

Miller pointed to the growing use of AI-generated search summaries and the rise of short-form video as key factors. Search engines now provide direct answers using generative AI instead of linking to external sources, while younger users increasingly turn to social video platforms rather than traditional websites.

Although Wikipedia’s knowledge continues to feed AI models, fewer people are reaching the original source.

The foundation warns that the shift poses risks to Wikipedia’s volunteer-driven ecosystem and donation-based model. With fewer visitors, fewer contributors may update content and fewer donors may provide financial support.

Miller urged AI companies and search engines to direct users back to the encyclopedia, ensuring both transparency and sustainability.

Wikipedia is responding by developing a new framework for content attribution and expanding efforts to reach new readers. The foundation also encourages users to support human-curated knowledge by citing original sources and recognising the people behind the information that powers AI systems.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Meta to pull all political ads in EU ahead of new transparency law

Meta Platforms has said it will stop selling and showing political, electoral and social issue advertisements across its services in the European Union from early October 2025. The decision follows the EU’s Transparency and Targeting of Political Advertising (TTPA) regulation coming into full effect on 10 October.

Under TTPA, platforms will be required to clearly label political ads, disclose the sponsor, the election or social issue at hand, the amounts paid, and how the ads are targeted. These obligations also include strict conditions on targeting and require explicit consent for certain data use.

Meta called the requirements ‘significant operational challenges and legal uncertainties’ and labelled parts of the new rules ‘unworkable’ for advertisers and platforms. It said that personalised ads are widely used for issue-based campaigns and that limiting them might restrict how people access political or social issue-related information.

The company joins Google, which made a similar move last year citing comparable concerns about TTPA compliance.

While political ads will be banned under paid formats, Meta says organic political content (e.g. users posting or sharing political views) will still be permitted.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Quebec man fined for using AI-generated evidence in court

A Quebec court has fined Jean Laprade C$5,000 (US$3,562) for submitting AI-generated content as part of his legal defence. Justice Luc Morin described the move as ‘highly reprehensible,’ warning that it could undermine the integrity of the judicial system.

The case concerned a dispute over a contract for three helicopters and an airplane in Guinea, where a clerical error awarded Laprade a more valuable aircraft than agreed. He resisted attempts by aviation companies to recover it, and a 2021 Paris arbitration ruling ordered him to pay C$2.7 million.

Laprade submitted fabricated AI-generated materials, including non-existent legal citations and inconsistent conclusions, in an attempt to strengthen his defence.

The judge emphasised that AI-generated information must be carefully controlled by humans, and the filing of legal documents remains a solemn responsibility. Morin acknowledged the growing influence of AI in courts but stressed the dangers of misuse.

While noting Laprade’s self-representation, the judge condemned his use of ‘hallucinated’ AI evidence and warned of future challenges from AI in courts.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Adult erotica tests OpenAI’s safety claims

OpenAI will loosen some ChatGPT rules, letting users make replies friendlier and allowing erotica for verified adults. Altman framed the shift as ‘treat adult users like adults’, tied to stricter age-gating. The move follows months of new guardrails against sycophancy and harmful dynamics.

The change arrives after reports of vulnerable users forming unhealthy attachments to earlier models. OpenAI has since launched GPT-5 with reduced sycophancy and behaviour routing, plus safeguards for minors and a mental-health council. Critics question whether evidence justifies loosening limits so soon.

Erotic role-play can boost engagement, raising concerns that at-risk users may stay online longer. Access will be restricted to verified adults via age prediction and, if contested, ID checks. That trade-off intensifies privacy tensions around document uploads and potential errors.

It is unclear whether permissive policies will extend to voice, image, or video features, or how regional laws will apply to them. OpenAI says it is not ‘usage-maxxing’ but balancing utility with safety. Observers note that ambitions to reach a billion users heighten moderation pressures.

Supporters cite overdue flexibility for consenting adults and more natural conversation. Opponents warn normalising intimate AI may outpace evidence on mental-health impacts. Age checks can fail, and vulnerable users may slip through without robust oversight.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Teen content on Instagram now guided by PG-13 standards

Instagram is aligning its Teen Accounts with PG-13 movie standards, aiming to ensure that users under 18 only see age-appropriate material. Teens will automatically be placed in a 13+ setting and will need parental permission to change it.

Parents who want tighter supervision can activate a new ‘Limited Content’ mode that filters out even more material and restricts comments and AI interactions.

The company reviewed its policies to match familiar parental guidelines, further limiting exposure to content with strong language, risky stunts, or references to substances. Teens will also be blocked from following accounts that share inappropriate content or contain suggestive names and bios.

Searches for sensitive terms such as ‘gore’ or ‘alcohol’ will no longer return results, and the same restrictions will extend to Explore, Reels, and AI chat experiences.

Instagram worked with thousands of parents worldwide to shape these policies, collecting more than three million content ratings to refine its protections. Surveys show strong parental support, with most saying the PG-13 system makes it easier to understand what their teens are likely to see online.

The updates begin rolling out in the US, UK, Australia, and Canada and will expand globally by the end of the year.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

New YouTube tools provide trusted health advice for teens

YouTube is introducing a new shelf of mental health and wellbeing content designed specifically for teenagers. The feature will provide age-appropriate, evidence-based videos covering topics such as depression, anxiety, ADHD, and eating disorders.

Content is created in collaboration with trusted organisations and creators, including Black Dog Institute, ReachOut Australia, and Dr Syl, to ensure it is both reliable and engaging.

The initiative will initially launch in Australia, with plans to expand to the US, the UK, and Canada. Videos are tailored to teens’ developmental stage, offering practical advice, coping strategies, and medically-informed guidance.

By providing credible information on a familiar platform, YouTube hopes to improve mental health literacy and reduce stigma among young users.

YouTube has implemented teen-specific safeguards for recommendations, content visibility, and advertising eligibility, making it easier for adolescents to explore their interests safely.

The company emphasises that the platform is committed to helping teens access trustworthy resources, while supporting their wellbeing in a digital environment increasingly filled with misinformation.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Virtual hosts and mass output shake up fragile podcast industry

AI is rapidly changing the podcast scene. Virtual hosts, no microphones or studios, are now producing content at a scale and cost that traditional podcasters find hard to match.

One of the pioneers in this trend is Inception Point AI, founded in 2023. With just eight people, the company produces around 3,000 podcast episodes per week, each costing about one dollar to make. With as few as twenty listens, an episode can be profitable.

Startups like ElevenLabs and Wondercraft have also entered the field, alongside companies leveraging Google’s Audio Overview. Many episodes are generated from documents, lectures, local data, anything that can be turned into a voice-narrated script. The tools are getting good at sounding natural.

Yet there is concern among indie podcasters and audio creators. The flood of inexpensive AI podcasts could saturate platforms, making it harder for smaller creators to attract listeners without big marketing budgets.

Another issue is disclosure: many AI-podcast platforms do note that content is AI-generated, but there is no universal requirement for clear labelling. Some believe listener expectations and trust may erode if distinction between human vs. synthetic voices becomes blurred.

As the output volume rises, so do questions about content quality, artistic originality, and how advertising revenues will be shared. The shift is real, but whether it will stifle creative diversity is still up for debate.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

EU nations back Danish plan to strengthen child protection online

EU countries have agreed to step up efforts to improve child protection online by supporting Denmark’s Jutland Declaration. The initiative, signed by 25 member states, focuses on strengthening existing EU rules that safeguard minors from harmful and illegal online content.

However, Denmark’s proposal to ban social media for children under 15 did not gain full backing, with several governments preferring other approaches.

The declaration highlights growing concern about young people’s exposure to inappropriate material and the addictive nature of online platforms.

It stresses the need for more reliable age verification tools and refers to the upcoming Digital Fairness Act as an opportunity to introduce such safeguards. Ministers argued that the same protections applied offline should exist online, where risks for minors remain significant.

Danish officials believe stronger measures are essential to address declining well-being among young users. Some EU countries, including Germany, Spain and Greece, expressed support for tighter protections but rejected outright bans, calling instead for balanced regulation.

Meanwhile, the European Commission has asked major platforms such as Snapchat, YouTube, Apple and Google to provide details about their age verification systems under the Digital Services Act.

These efforts form part of a broader EU drive to ensure a safer digital environment for children, as investigations into online platforms continue across Europe.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!