News · Loktech

2026-05-07 Global

Global report: AI use grows, but the North-South gap widens

Microsoft’s 2026 AI Diffusion Report finds 17.8 percent of the world’s working-age people now use AI, with the Global South falling further behind.

Microsoft’s chief data scientist Juan Lavista Ferres published the company’s Global AI Diffusion Report in May 2026, measuring what share of working-age people in each country actually use generative AI tools.

Usage rose to 17.8 percent worldwide in early 2026, with the United Arab Emirates leading at around 70 percent. But the divide is stark: about 27.5 percent of people in the Global North use AI against 15.4 percent in the Global South, and the gap widened during the quarter.

One hopeful signal: better AI support for Asian languages measurably increased adoption in countries like Japan, South Korea and Thailand, evidence that language support, not just infrastructure, decides who gets to use AI. For communities whose languages AI does not speak, the report quantifies the exclusion.

Source: Microsoft On the Issues

2026-03-25 USA

US lawmakers propose a freeze on giant AI data centres

Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced a bill to pause construction of large AI data centres until national safeguards exist.

Senator Bernie Sanders introduced the Artificial Intelligence Data Center Moratorium Act in the US Senate on 25 March 2026, with Representative Alexandria Ocasio-Cortez bringing the House version in June and several members of Congress joining as co-sponsors.

The bill would halt construction or expansion of data centres drawing 20 megawatts or more of power until Congress passes protections covering electricity prices, water use, privacy and community impacts. It follows complaints across the US that AI facilities raise household power bills and strain local water supplies.

More than a hundred local communities in the US have already passed their own data centre moratoriums. Whether or not the bill becomes law, it signals that the physical footprint of AI, the land, power and water behind the chat window, has become a mainstream political issue.

Source: Office of Senator Bernie Sanders

2026-02-26 Global

Michael Running Wolf: teaching AI to serve endangered Indigenous languages

A Northern Cheyenne computer scientist is building speech technology and youth coding camps so Indigenous languages survive the AI era on their own terms.

Michael Running Wolf, a Northern Cheyenne computer scientist raised in rural Montana, once worked as an engineer on Amazon’s Alexa. He left to ask a different question: why should voice assistants speak every major language yet not one of the hundreds of Indigenous languages of the Americas?

He co-founded First Languages AI Reality with the Mila AI institute in Montreal to build speech recognition for Indigenous languages, whose complex grammar defeats standard tools. With his wife Caroline he also runs Lakota AI Code Camps, where Native teenagers learn to build AI on their own data.

Reporting in February 2026 profiled his insistence on data sovereignty: communities keep ownership of their recordings rather than donating them to technology companies. His argument is that language revival and AI skills can reinforce each other, provided the community holds the keys.

Source: Prism Reports

2026-02-21 India

India hosts the AI Impact Summit; 92 countries adopt the New Delhi Declaration

The largest global AI gathering yet was held in New Delhi in February 2026 and closed with a declaration on sharing AI’s benefits equitably.

India hosted the AI Impact Summit in New Delhi from 16 to 21 February 2026, the successor to earlier AI summits held in the UK, South Korea and France, with hundreds of thousands attending in person and online.

The closing declaration, endorsed by 92 countries and international organisations, was framed by the phrase Sarvajan Hitaya, Sarvajan Sukhaya, welfare and happiness for all. Initiatives included a Global AI Impact Commons of shareable use cases from over 30 countries and a playbook with the International Labour Organization for preparing workers.

More than 200 billion US dollars of AI investment commitments were announced around the summit. For India, the event marked a claim to speak for the Global South in how AI is governed; what the declaration changes in practice will depend on follow-through.

Source: Press Information Bureau, Government of India

2026-02-19 India

Delhi: an AI voice clone of a daughter used to steal 2 lakh rupees from her mother

Fraudsters cloned a daughter’s voice from social media clips and staged a fake kidnapping call to a 65-year-old woman in Delhi.

A 65-year-old woman in Delhi received a call in February 2026 in which a sobbing voice, identical to her daughter’s, begged to be saved from kidnappers. Under pressure to pay instantly, she transferred 2 lakh rupees. Her daughter had been safe at home the whole time.

Investigators say fraudsters harvest short voice clips from social media; a few seconds of clear audio are enough for AI tools to build a convincing clone. Callers add crying and background noise and demand immediate UPI or bank transfers, leaving no time to verify.

Police and cybersecurity groups expect such AI-enabled fraud to rise steeply and advise families to agree on a code word, to call back on the known number, and to treat any urgent money demand as suspicious; in India, fraud should be reported immediately to the 1930 helpline. Elderly people living alone are the most targeted.

Source: The420.in

2026-01-15 India

Bhashini: donating your voice to build Indian-language AI

India's initiative lets people donate their voice to build community language-data for AI.

Bhashini is run by the Ministry of Electronics and IT as India’s National Language Translation Mission. It was launched by Prime Minister Narendra Modi in July 2022 with a clear aim: the internet, and now AI, should work in all 22 scheduled Indian languages, not only in English and Hindi.

Its crowdsourcing arm, Bhasha Daan, invites ordinary citizens to donate their voice by reading sentences aloud, to type what they hear, and to check translations made by others. These contributions become open datasets and models that any developer can build upon.

Bhashini tools now sit behind translation in government services and apps, and the project is discussed internationally as a model of a state building public language infrastructure rather than leaving it to private companies. For speakers of smaller languages, it is one of the few official doors into the AI world.

Source: Digital India Bhashini Division

2025-08-01 Africa

Masakhane: community-built translation for 30+ African languages

African communities built their own translation tools, a model example of participatory AI.

Masakhane means ‘we build together’ in isiZulu. The initiative began around 2019, when African natural language processing researchers, among them Jade Abbott, a software engineer from South Africa, grew tired of African languages being studied mostly by outsiders, or not at all.

It runs as an open, volunteer collective: anyone can join its channels, contribute data or code, and co-author research. Members across more than 30 African countries have built translation models and datasets for dozens of languages and published together at leading conferences.

Masakhane’s way of working, research by African communities for African languages with shared authorship, has become a template cited by language communities worldwide, and several of its members have gone on to shape language AI work at universities and labs.

Source: Masakhane

2025-06-10 Global

CARE principles for Indigenous data governance gain ground

Communities' right to control and benefit from their own data is gaining traction worldwide.

The CARE principles (Collective benefit, Authority to control, Responsibility, Ethics) were drafted by the Global Indigenous Data Alliance, a network of Indigenous researchers and data specialists including Stephanie Russo Carroll (Ahtna) and Maui Hudson (Maori).

They answer a gap in the open-data movement: the older FAIR principles say data should be findable and reusable, but say nothing about who governs it. CARE adds that Indigenous peoples must control data about themselves, their lands and their languages, and must benefit from its use.

Universities, archives and funding bodies in several countries now reference CARE in their data policies, and the principles are a key reference point in debates about who may collect language recordings and build AI from them. For any community handing over its voice, they are a checklist worth knowing.

Source: Global Indigenous Data Alliance

2024-07-01 Global

Mozilla Common Voice: a public voice dataset anyone can add their language to

Volunteers around the world read sentences aloud to build an open dataset of voices in over 100 languages, so that speech technology is not limited to the biggest languages.

Common Voice was started in 2017 by Mozilla, the nonprofit behind the Firefox browser, because speech datasets were locked inside a few large companies and covered only a handful of major languages.

The mechanics are simple: volunteers read sentences aloud on the website, and other volunteers listen and confirm the recordings are correct. Everything is released into the public domain, so anyone, from a student to a startup, can train speech tools with it.

Communities have used it to put their languages on the map: Kinyarwanda speakers in Rwanda, for example, built one of the largest voice datasets in the collection, which powered local voice assistants. It is working proof that ordinary speakers can create the raw material of speech AI themselves.

Source: Mozilla Foundation

2023-09-05 Africa

Deep Learning Indaba: strengthening African machine learning

An annual gathering builds a community of African researchers so that machine learning in Africa is shaped by Africans, with local events across dozens of countries.

The Deep Learning Indaba was founded in 2017 by African researchers, among them Shakir Mohamed, Ulrich Paquet and Vukosi Marivate, who counted how few African authors appeared at the world’s main machine learning conferences and decided to change it from the inside.

‘Indaba’ is an isiZulu word for a gathering to discuss important matters. The annual meeting moves between African countries and is paired with IndabaX, smaller locally organised events that have spread to dozens of countries, along with mentorship programmes and awards for African research.

The community has grown into thousands of researchers and engineers, and its alumni now work in universities, startups and AI labs worldwide. It is a long-term answer to the question of who gets to build AI for Africa: Africans themselves.

Source: Deep Learning Indaba

2023-08-01 India

Karya: rural Indians earn fair wages creating language data

A nonprofit platform pays rural workers in India well above market rates to record speech and text in their own languages, and gives them ownership and royalties when the data is resold.

Karya grew out of research at Microsoft Research India and became a nonprofit in 2021, co-founded by Manu Chopra, a computer scientist who grew up in Delhi and studied at Stanford. Its idea: data work for AI can pay a dignified wage instead of an exploitative one.

Rural workers, many of them women, record speech and transcribe text in their own languages on their phones. Karya pays many times the usual market rate for such work, and, unusually, workers keep a stake in the data: when a dataset is sold again, royalties flow back to the people who created it.

Its speech data in Indian languages now feeds major AI projects, and Chopra was included in TIME’s first list of influential people in AI. Karya is regularly cited as proof that the supply chain of AI does not have to run on underpaid labour.

Source: Karya

2023-07-07 Australia

Australia: the Robodebt scheme issued unlawful automated debt notices

An automated income-averaging system wrongly told hundreds of thousands of welfare recipients that they owed money; a Royal Commission found the scheme unlawful and recorded its heavy human cost.

Robodebt was the nickname of an automated debt recovery scheme run by Australia’s welfare agency Centrelink from 2015 to 2019. It replaced human checks with a simple calculation that averaged a person’s yearly income across fortnights, and assumed any mismatch meant cheating.

Around 470,000 debts raised this way were legally invalid, and the burden of proof was pushed onto welfare recipients, many of whom paid money they never owed. Families gave evidence linking the stress of the false debts to breakdowns and suicides of loved ones.

A class action forced roughly 1.8 billion Australian dollars in repayments and settlements, and a Royal Commission led by former judge Catherine Holmes found the scheme unlawful and referred officials for investigation. Robodebt is now taught worldwide as a warning about automating welfare decisions.

Source: Royal Commission into the Robodebt Scheme

2023-05-25 India

AI4Bharat: open models and datasets for Indian languages

A research centre at IIT Madras releases open speech and translation models and datasets for 22 scheduled Indian languages, built with contributions from speakers across the country.

AI4Bharat is a research centre at IIT Madras led by professors Mitesh Khapra and Pratyush Kumar, with support from the philanthropy of Nandan Nilekani. Its mission is open AI infrastructure for the 22 scheduled languages of India.

Teams have travelled to districts across the country to record everyday speech from farmers, shopkeepers and homemakers, so that models hear real voices and dialects rather than studio recordings. The centre releases its models and datasets, such as the IndicTrans translation family, openly.

Its systems are used inside public projects including Bhashini, and its datasets have become standard building blocks for anyone making Indian-language AI. It shows what representation takes for languages like those of Marwar: deliberate, funded, field-level data collection.

Source: AI4Bharat, IIT Madras

2023-01-18 Africa

Kenya: workers paid under 2 dollars an hour to filter toxic content for ChatGPT

A TIME investigation found that Kenyan workers were paid less than 2 dollars an hour to read and label violent and abusive text so that ChatGPT could avoid producing it; many were left with lasting trauma.

Behind ChatGPT’s politeness lies filtering work done by people. TIME reporter Billy Perrigo revealed that OpenAI had contracted Sama, an outsourcing firm with offices in Nairobi, to have Kenyan workers label extremely violent and abusive text so a safety filter could learn to block it.

The workers, paid roughly 1.32 to 2 US dollars an hour depending on role and performance, read disturbing material for hours each day. Several, including Mophat Okinyi, later described lasting psychological damage and strained family lives; counselling support was described as inadequate.

After the story, the contract was ended early, workers helped form a content moderators’ union in Nairobi, and petitions were filed with Kenya’s parliament. The case put the hidden human labour of AI, and where in the world it is placed, permanently on the public agenda.

Source: TIME

2021-10-25 Europe

Netherlands: a childcare benefits algorithm ruined thousands of families

Dutch tax authorities used an algorithm that treated people with dual nationality as fraud risks; tens of thousands of families were wrongly made to repay benefits, and the government resigned over the scandal.

In the Netherlands, the tax office used a risk-scoring system to hunt for fraud in childcare benefit applications. Holding a second nationality was treated as a risk signal, which meant families of Moroccan, Turkish, Surinamese and other backgrounds were flagged far more often.

Tens of thousands of parents were falsely branded fraudsters and ordered to repay their entire benefit, often tens of thousands of euros, with no explanation and no effective appeal. Families lost homes and marriages, and more than a thousand children ended up placed in care.

The scandal forced Prime Minister Mark Rutte’s entire cabinet to resign in January 2021. Amnesty International’s report Xenophobic Machines documented the discrimination, the data authority fined the tax office, and the Netherlands created a public register of government algorithms in response.

Source: Amnesty International

2021-03-01 New Zealand

Te Hiku Media: Maori communities build their own speech recognition

A Maori media organisation collected thousands of hours of te reo Maori speech with community consent and built its own speech recognition, keeping the data under tribal guardianship instead of handing it to big technology companies.

Te Hiku Media is a small Maori radio station in the far north of New Zealand, led by Peter-Lucas Jones and engineer Keoni Mahelona. Decades of its archived broadcasts hold some of the best recordings of fluent te reo Maori speakers.

In 2018 it ran a campaign in which community members read phrases aloud, gathering hundreds of hours of labelled speech within days. With that, its own engineers trained a speech recognition system for te reo Maori that reached accuracy comparable to commercial tools.

The organisation released its work under a kaitiakitanga (guardianship) licence, which keeps control with the Maori community, and it has publicly refused requests from large technology companies to hand the data over. It has become the most cited example of Indigenous data sovereignty in AI.

Source: Te Hiku Media

2020-08-17 UK

UK: an algorithm downgraded students exam results

When exams were cancelled in 2020, an algorithm estimated grades partly from schools past results, downgrading nearly 40 percent of teacher assessments and hitting students from poorer schools hardest; it was scrapped after protests.

When COVID-19 cancelled school-leaving exams in 2020, England’s exam regulator Ofqual decided grades would be produced by a formula that combined teachers’ rankings of students with each school’s results in previous years.

On results day, nearly 40 percent of teacher-assessed A-level grades were pushed down. Because the formula leaned on school history, bright students from large, poorer state schools were downgraded most, while small classes at private schools gained; students protested outside the education department against ‘the algorithm’.

Within days the government scrapped the results and reverted to teacher grades, and the head of Ofqual left her post. The episode became a standard example of how an apparently neutral formula can freeze past inequality into individual futures.

Source: BBC News

2020-06-24 USA

Detroit: a man was arrested because face recognition matched the wrong person

Robert Williams was held for 30 hours after face recognition software wrongly matched his driving licence photo to shop CCTV footage; several similar wrongful arrests have followed, mostly of Black men.

Robert Williams, an office worker from a Detroit suburb, was arrested on his front lawn in front of his wife and young daughters in January 2020, for a watch-shop theft he had nothing to do with.

Detroit police had run a blurry CCTV image through face recognition software, which returned his old driving licence photo as a match. Investigators treated the machine’s suggestion as fact; Williams was held for about 30 hours before the case fell apart.

With the ACLU he sued the city, and the 2024 settlement forces Detroit police to stop arresting anyone on a face-recognition match alone. Several similar wrongful arrests, nearly all of Black men and women, have since been documented across the United States.

Source: The New York Times

2018-02-01 USA

Gender Shades: face analysis fails darker-skinned women the most

MIT researchers tested commercial face analysis systems and found error rates of up to 35 percent for darker-skinned women, against under 1 percent for lighter-skinned men.

Joy Buolamwini, a Ghanaian-American researcher at the MIT Media Lab who founded the Algorithmic Justice League, teamed up with Timnit Gebru to ask a simple question: does commercial face analysis work equally well for everyone?

They built a balanced test set of faces from the parliaments of three African and three European countries and ran systems from IBM, Microsoft and Face++ over it. Error rates for darker-skinned women reached almost 35 percent; for lighter-skinned men they stayed under 1 percent.

The companies rushed to improve their products, IBM later withdrew from face recognition entirely, and the study armed lawmakers pressing for limits on the technology. Gender Shades made ‘whose faces was this trained on?’ a question that no longer sounds academic.

Source: MIT Media Lab

2016-05-23 USA

Machine Bias: a US risk score marked Black defendants as future criminals

ProPublica investigated a risk-scoring tool used in US courts and found it wrongly labelled Black defendants as high risk at nearly twice the rate of white defendants.

In 2016, journalists at ProPublica, led by Julia Angwin, obtained the risk scores that a commercial tool called COMPAS had assigned to more than 7,000 people arrested in Broward County, Florida, and then checked who actually reoffended in the following two years.

The scores were wrong in opposite directions for different groups: Black defendants who did not reoffend were nearly twice as likely to have been labelled high risk, while white defendants who did reoffend were more often labelled low risk. Judges saw these scores when deciding on bail and sentences.

The maker disputed the analysis, and the argument over what fair even means mathematically launched an entire research field on algorithmic fairness. Machine Bias remains the reference case for hidden discrimination inside a score.

Source: ProPublica

Critical & Participatory AI News