Month: January 2026

India’s Proposed Compulsory Licensing Scheme for AI Training: It Could be the Worst of All Worlds.

A wooden gavel striking a golden copyright symbol with the text 'COMPULSORY LICENSING' in bold.

Used with permission

In early December the Indian government through the Department for Promotion of Industry and Internal Trade (DPIIT), launched a public consultation on its proposal for a “One Nation, One Licence, One Payment” regime to govern AI training on copyrighted content. The comment period closes early next month. The stated intent is to establish a framework to allow AI developers to access proprietary content for AI algorithm training, without rightsholder authorization, while ostensibly taking into account the concerns and economic interests of those same rightsholders. The proposal from a semi-official committee established by DPIIT is based on a compulsory licence regime covering all content, past, present and future, with no opt outs for content owners. Given the unrepresentative nature of the committee, which failed to include rightsholders from many creative sectors, it is not surprising it has managed to come up with a solution that pleases absolutely no-one, not rightsholders nor the AI industry. Moreover, it is probably unworkable and could rightly be described as the “worst of all worlds”. How did we end up here?

It is widely accepted that to further refine and develop AI, more data is constantly required. Quality data is important. The better the data, the better the product. Many AI developers have, without authorization, already helped themselves to vast amounts of copyright protected material either through accessing pirate databases or by simply hoovering up publicly accessible (but nonetheless copyright protected) content on the internet. The result has been a series of lawsuits launched by rightsholders, primarily in the US but also in India, the UK and Canada. In the US the parameters of the fair use doctrine are being tested, with mixed results. There have been some settlements, notably the Bartz v Anthropic case in which AI developer Anthropic agreed to pay a group of authors $1.5 billion for having accessed and reproduced their works through pirate websites (and without authorization), as well as a large number of voluntary licensing agreements between corporate rightsholders such as media outlets, publishers and music labels, and AI companies. Clearly, legal leverage is needed to convince AI developers to negotiate with rightsholders (no one is volunteering, that’s for sure). In the US this is backed up by a statutory damages’ regime where the risk of being hit with large statutory damages per infraction is a strong incentive for the AI industry to reach licensing deals. Unfortunately, many countries do not have a statutory damages provision in their copyright law, and India is one of them.

Ironically, outside the US where the fair use case-by-case legal doctrine does not exist, AI developers are arguably in even greater legal jeopardy. As a result, they have pushed for a wide statutory text and data mining (TDM) exception to be introduced into national copyright laws. The AI industry is doing this in India, although there is strong opposition to introducing a TDM escape hatch. When a TDM exception applies, rightsholders are paid nothing. The DPIIT “solution” proposes to address the non-payment issue–but is going about it in precisely the wrong way. Appropriating the IP of all rightsholders by imposing a draconian compulsory license regime penalizes rather than rewards rightsholders and imposes a lowest-common-denominator value on all content, not distinguishing between premium and pedestrian content while taking away from rightsholders the inherent right to determine how and where their content is used.

The compulsory licence regime proposed by DPIIT would be administered by a new non-profit body, a Collective Management Organization (CMO) to be established by statute. This would create another layer of bureaucracy harking back to the days of the “licence raj”. A government appointed committee would set rates in conjunction with the new non-profit. Royalties would be applied retroactively as well as prospectively. Existing voluntary licensing agreements between content providers and AI companies would have to be terminated and future licence agreements prohibited. Rightsholders who do not want to licence their content could not opt-out. This is overreach at its worst.

The tech industry represented by NASSCOM, the National Association of Software and Service Companies, that also includes Google and Microsoft, vehemently opposes the One Nation, One Licence, One Payment proposal, arguing instead for a TDM exception. It also argues that rightsholders who wish to opt out should be able to do so, thus avoiding AI companies having to pay royalties for content it does not want or has already licensed. It also doesn’t want the precedent of compulsory payments, as India would be the first country globally to institute such a regime. NASSCOM claims that a compulsory licence regime would slow innovation, the card always played by the AI industry when the issue of protecting rightsholders is discussed. However, a TDM exception is not currently on offer in India. It is a “solution” that is facing increasing pushback in many countries. Australia has just rejected the concept, and other jurisdictions are examining it critically.  Where a TDM exception does exist, as in the UK and EU, its use is constrained and subject to several conditions, such as in Britain where it is limited to non-commercial research.

Not only is the tech industry strongly opposed to the compulsory licence proposal, so are rightsholders such as the broadcasting industry and Bollywood. Not only would a compulsory licence be extremely difficult to implement given the nature of CMOs which are generally highly inefficient and administratively cost heavy– not to mention that the Committee’s working paper proposes two layers of CMO (thus double handling and processing)– but the draconian and sweeping nature of the compulsory licence is of great concern. Rightsholders are given no ability to opt out or to refuse to have their content conscripted. In fact, the proposal includes a provision granting AI developers access to content as a “matter of right”. What happened to the fundamental right of an author to determine how, when and even if their content is to be reproduced? Compulsory licences are extraordinarily blunt instruments and do not work when sophisticated content like audio-visual and music products are involved. There are many elements to licensing, including contractual issues that go beyond price that are carefully negotiated and carry with them specific obligations and privileges. A compulsory licence is a “one size fits all” solution. It strips away the rights of content owners and is, in effect, a form of “compensated expropriation”. And the compensation is minimal.

Having pleased no-one with its proposal, it remains to be seen where DPIIT will go next once all comments are received and evaluated. A follow-up proposal is expected that will deal with outputs, just as the initial one did with inputs. India, like many countries, wants to participate in the global development of AI. Its rich local-language content is a strategic asset. Imposing a compulsory licence would stifle the creativity that drives this cultural comparative advantage. In the absence of a market failure there is no rationale for resorting to the sledgehammer of a compulsory licence for all content. The preferred solution for rightsholders, voluntary licensing, is also becoming a preferred solution for AI developers as the legal ground on which they are operating in accessing content without authorization is looking increasingly shaky. Voluntary licensing is growing globally as AI developers absorb this reality. If wide TDM exceptions are off the table, AI developers will have no recourse but to negotiate licence agreements. (A statutory damages regime would provide even greater incentive to do so). Bringing in a wide TDM exception or worse, introducing a blunt and highly bureaucratic compulsory licence regime, is a sure way to kill the growing voluntary licensing market.

DPIIT’s next steps will be crucial. If the goal is to create the conditions for a “made in India” AI industry while nurturing and protecting India’s valuable cultural assets there is no better solution than to create the conditions for the growth of a voluntary licensing market. Strong cultural industries and robust AI development go hand-in-hand. This means dispensing with forced, bureaucratic solutions like the proposed compulsory licensing regime while holding firm on rejecting a TDM loophole that would allow AI companies to plunder India’s cultural richness without any compensation to creators and rightsholders. Let’s hope India gets it right.

© Hugh Stephens, 2026.  All Rights Reserved

The Online Streaming Act or Dairy Supply Management: Which one should Canada Surrender to the US in CUSMA Trade Negotiations? Or is it a Question of Putting Some Water into the Wine of Both?

A person pouring water from a ceramic jug into a large clay pot.

Image: Shutterstock

Should Canada give up the Online Streaming Act (OSA) in forthcoming CUSMA negotiations in order to preserve dairy supply management, as a former Vice Chair of the CRTC, Peter Menzies, suggested in a Globe and Mail oped earlier this month? Perhaps he was just being deliberately provocative although the question hits one of the raw nerves of Canadian politics. The cultural community– particularly in Quebec–would be enraged if this happened. But then if supply management is watered down to allow more imports from the US, especially in dairy, the dairy farmers–particularly in Quebec–will be equally enraged. Which group has the greater political clout? In both cases, Quebec-based interest groups have a card to play denied to others in Canada. It is called the Bloc Quebecois, and if enough support bleeds from the Liberals to the Bloc, that could just open the way to the Conservatives to form the national government they so desperately crave. The cultural mavens in Toronto have little choice; either support the Liberals or face a worse fate when those Conservative cowboys from Alberta take the reins of power.

The Quebec cultural community which insists that measures are needed to ensure that foreign streamers both contribute financially to support Canadian content (Cancon) and ensure that Cancon (when expressed in French) is “discoverable”, has yet another card up its sleeves. It is called Bill 109, Quebec legislation (that is probably ultra vires since broadcasting clearly falls within federal jurisdiction) that purports to regulate the discoverability of French-language cultural content in the digital environment. If Canada gives way on the Online Streaming Act in CUSMA negotiations, watch Quebec fill the void. So where does all this leave the Carney government? Between a rock and a hard place.

It is true, as Menzies has pointed out, that the CRTC has been very slow, plodding even, in dealing with implementation of the OSA. It may even be overwhelmed, with inadequate staff as he suggests. The fact that implementation is still a work in progress makes it easier for the US government to bring pressure to stop or at least to modify rollout of the legislation, whereas other objectives mentioned in recent USTR hearings, such as changes to the Bank Act to benefit US financial institutions or measures to terminate supply management would require significant legislative and regulatory change to undo measures that have been in place for decades. Best to nip it in the bud, or to kill it in the egg, as they say in Quebec.  

Canada’s planned introduction of a Digital Services Tax (DST) is a prime example of a nipped-in-the-bud policy. A DST deals with tax avoidance measures implemented by large digital multinationals by taxing their in-country revenues rather than their manipulated profits. Some countries, such as the UK, France, Spain, Italy etc had already implemented a DST before Trump’s return to office and seem to have got away with it, even though Google, Microsoft, Amazon, META and others of that ilk have the Trump Administration’s ear. Canada intended to implement a DST several years ago but dithered and dragged its feet, finally passing legislation in 2024 that would have brought a DST into effect on June 30, 2025, backdated to 2022 when the law should have been put into effect in the first place. Unfortunately for Canada, the implementation date fell right in the middle of the Trump tariff war and Canadian efforts to negotiate some relief. But rather than postponing implementation yet again–and using the possibility of a future DST as negotiating collateral–Canada “bravely” announced it was going ahead with implementation (regardless of the consequences). Until it wasn’t. Trump tweeted that he was cancelling trade negotiations with Canada because of the DST and voilà, over a weekend, the DST was cancelled (on June 29, 2025).

Trade talks resumed and actually appeared to be making some progress with respect to sectoral tariffs such as steel until the next excuse Trump found to end them. This time it was over Ontario’s World Series free trade ads that ran on US television, using Ronald Reagan’s words from a 1980s era speech praising free trade and condemning protectionist tariffs. The content of those ads may have been accurate, but the result was one of Canada’s more prominent “own goal” moments. While Doug Ford may have derived some brief satisfaction from getting under Donald Trump’s skin, the steelworkers of Sault Ste. Marie, who might have benefited from a rumoured sectoral deal on steel, have been paying the price. I think this fiasco helps explain the public anger of US Ambassador to Canada Pete Hoekstra (who surely wins the 2025 “Bull in a China shop” award) who crudely vented his frustration that a deal so close to fruition got blown out of the water through Premier Doug’s ill-considered initiative.

But what about supply management? Canada should be taking a long, hard look at the wisdom of continuing to defend this 1970s policy that almost every other country has since abandoned. Instead, it should use the CUSMA negotiations as the reason to ditch a monopoly that protects a few chosen producers of supply managed commodities at the expense of consumers and the rest of the economy. Unfortunately, that won’t happen because of Canadian political realities but there is still scope for some wiggle room. In recent years, Canada has been forced as part of its trade negotiations to open slivers of the dairy market to EU countries, CPTPP trading partners, and to the US through the CUSMA. The dairy industry screamed blue murder but was paid off for having to face a bit more competition. As part of liberalizing as little as possible, Canada routinely plays games with its commitments by awarding import quotas to the same domestic dairy industry with which exporters of dairy products to Canada are competing. Some additional foreign cheese and dairy products become available to consumers but in effect the fox is in charge of deciding which chickens get let in, and at what price. Even though this policy is an albatross around Canada’s neck, such is the power of the dairy industry (which is reputed to control the outcome of no less than eight ridings in Quebec) that all political parties support keeping supply management off the table in all trade negotiations, and passed legislation to this effect. In a political environment where the government is one vote short of a majority, risking the ire of Quebec dairy farmers is a risky business.

Does that mean that supply management is completely off the table and instead there should be another sacrificial lamb, such as domestic broadcasting and cultural policy, as Peter Menzies has suggested? This is a doubtful proposition. Despite all the posturing about supply management being “off the table”, there will almost certainly be some concessions to the US, even if it is only in the way the tariff free import quotas are managed. The Carney government will claim it is defending supply management, while making some tweaks to the system. It can do the same for cultural industries. Defend the essence but find compromises that US industry can live with.

Like supply management, the Online Streaming Act also has wiggle room in its implementation. Already we have seen the CRTC announce changes to Cancon definitions that introduce greater flexibility and go some way toward meeting the concerns of the (largely US-based) content streamers, while preserving elements of protection for Canadian production. (Canadian makeup and hair design artists will be happy as use of their services adds an element of “Canadianness” to a production that could be useful in meeting the Cancon definition. This just goes to show that you can never discount the influence of a specific lobby). While the US has laid out some maximum wish-list objectives, including withdrawal of the Online Streaming Act (as well as the Online News Act), there are domestic political realities in Canada that will constrain Canadian trade negotiators from sacrificing the cultural sector to gain other objectives, just as there are with regard to supply management. The US may hold a big stick in the negotiations, but Canada is not without cards to play. It just has to be careful how to play them, and when mobilizing support inside the US to do so in a way that does not offend the touchy amour-propre of Donald Trump.

The end result for Canada will not be water or wine, but rather how much water to allow into the wine. Some dilution will be necessary but at the end of the day, for domestic political reasons (particularly in Quebec), the liquid in the glass still will still have to resemble wine more than water. This applies equally to cultural industries and broadcasting as well as to supply management. It is far from an either/or situation.

© Hugh Stephens, 2026. All Rights Reserved.

Vietnam’s New IP Law: The General Direction Makes Sense but Regulatory Details will be Key

Three musicians perform on stage at night, surrounded by colorful decorations and ambient lights.

Image: Shutterstock

On December 10, Vietnam’s National Assembly adopted an amended IP (intellectual property) law designed to update the legislation for the digital age, and especially to address issues involving the intersection of AI and IP. The next phase will be drafting of the regulations necessary to bring the law into effect by April 1, 2026. The law does several things; it will explicitly recognize IP as an economic asset allowing it to be used for financial and accounting purposes, and as collateral for loans; it clarifies that works autonomously generated by AI without substantive human creative input are not eligible for IP protection; and it allows the use of lawfully published works and data for AI research, testing, and training, provided that such use does not unreasonably prejudice the legitimate interests of rights holders and the resulting outputs do not infringe copyright. While all this sounds good on the surface, there are potential hidden problems with the AI research, testing and training provisions. In drafting the implementing regulations, Vietnam needs to be careful to not undermine its vibrant creative industries that depend on copyright protection to thrive.

The translated wording of the AI provision (Clause 5, Article 7) is rendered in English as follows:

“Organizations and individuals may use text and data related to intellectual property rights that have been lawfully published and made publicly accessible for purposes of scientific research, experimentation, training of artificial intelligence systems, provided that such use does not unreasonably affect the legitimate rights and interests of the author or the holder of intellectual property rights as stipulated by this Law.

For text and data that are subject to copyright and related rights protection, the use of such text and data under this Clause must also comply with the Government’s regulations.”

At first glance, this looks like another “standard” Text and Data Mining (TDM) exception, a carve out from copyright protection being pushed by the AI industry in a number of Asian countries, as I recently wrote about here. In that blog post, I noted that lawmakers and regulators in Asia are grappling with a common problem; how to incentivize the development of responsible AI while continuing to encourage and promote all-important content industries. Vietnam is facing the same issue. Throwing the cultural sector under the bus in the hopes of attracting some ephemeral hi-tech AI jobs is a false bargain. Strong cultural industries enable the development of strong content licensing markets for AI development, enabling a virtuous circle of further creativity. A strong cultural sector and strong, sustainable digital industries, especially those powered by AI, go hand-in-hand. To this end, Vietnam has put in some important guardrails, such as requiring that any copyright protected works used for AI training (1) be lawfully accessed and (2) be made publicly accessible for the purposes of training of artificial intelligence systems (among other uses). Additionally, language has been adopted from the Berne Convention “three step test” that requires the use to “not unreasonably affect the legitimate rights and interests of the author”. So far so good, but these features may not be adequate to protect the interests of Vietnam’s creative industries unless made explicit. This is where careful drafting of the regulations comes into play.

First, on lawful access, this is positive in that it rules out the use of pirated content to train AI systems (if enforced) and should make it illegal to bypass a paywall (digital lock, or technological protection method, TPM) to access content. However, this provision provides only relatively thin protection since lawful access can be obtained by purchasing the cheapest or lowest minimal access possible and then using that access to justify commercial exploitation. For example, an AI company could purchase exactly one copy of each book that it wants to ingest, or one copy of a sound or audio-visual recording.

With respect to the requirement that content be made “publicly accessible” before it can be used for training, this implies that any content not made publicly accessible should be off limits to AI developers. Therefore, disclaimers or terms of service specifying what uses are acceptable must be respected along with technical indicators, such as robots.txt protocols, that mark content that is not to be copied. Again, regulation will need to specifically deal with these points to avoid any ambiguity. On the basis of the “trust but verify” principle, AI developers should be subject to maximum transparency requirements and must be required to document all protected content used for training.

To give effect to the language taken from the Berne Convention requiring the use to not unreasonably affect the legitimate rights and interests of the author, this could also be addressed by regulation to make the meaning precise. For example, AI outputs that substitute for the originals on which they were trained clearly prejudice the rights and interests of the rights-holder. It is not enough that an AI produced output does not directly mimic the original; if it displaces it in the market, it is also damaging the economic interests of the author.

Finally, the law indicates that the use of text and data subject to copyright “must also comply with the Government’s regulations.” It is far from clear what this means. The use of text and data already must comply with the terms of the legislation as described above. This phrase would hardly seem necessary unless it is intended to provide some sort of general override that would invalidate the guardrails embedded in the law. This should be clarified in regulation or dropped entirely.

Vietnam has a lively and thriving cultural sector that is a key manifestation of its cultural expression and sovereignty. The Vietnam Creator’s Coalition has estimated that the arts, entertainment and recreation sector employed over 281,000 people and contributed VND 55,694 billion (USD $2.12 billion) to the Vietnamese economy. The cornerstone of this economic strength and growth is Vietnam’s copyright framework. The National Target Program for Cultural Development has set a target for the cultural industries to contribute 7 percent of the country’s GDP by 2030. If copyright protection is inadvertently gutted by poor application of the new law, those targets will be unattainable and the sector, an important instrument of national expression, will suffer.

The other element of the content protection environment in Vietnam that continuously undermines both domestic and international rightsholders is piracy. Vietnam is a known centre of audiovisual piracy, hosting numerous sites that operate more or less openly. It has been called out by USTR’s Special 301 report on many occasions, most recently again in 2025. Initially the worst offender was a pirate operation going by the name of 123Movies. More recently it has been Phimmoi and FMovies. Enforcement is weak and when a pirate site is shut down, usually after prolonged effort by rightsholders, it is not long before it is back in business. It is not just international rightsholders who suffer. The mass copying of the film “Red Rain” (Mưa Đỏ), the 2025 Vietnamese epic historical war drama adapted from a novel about the Second Battle of Quảng Trị in 1972, a fierce and protracted battle in the Vietnam War, is Exhibit A for the damage wrought by domestic piracy, as outlined by the British legal firm Rouse. The inability or unwillingness of the Vietnamese authorities to take effective action against openly operating pirate sites is a blot on Vietnam’s IP record and undermines the credibility of the stated objectives of the new IP law.

From the overall thrust of the law, it seems that Vietnam’s lawmakers want to protect the country’s valuable cultural assets and creative energy while providing scope for AI developers to have access to protected content under clearly defined conditions that respect the legitimate rights of creators. The regulations that will be drafted over the next few months need to make this explicit and apparent or else the intent of the law could be subverted. Robust action is also required to end the free ride being afforded to content pirates. It is often said that the proof of the pudding lies in the eating. In this case, the proof of the legislation lies in the drafting of sound regulations to implement the new IP law effectively, while taking meaningful action to curb the scourge of AV piracy.

© Hugh Stephens, 2026. All Rights Reserved.

Deloitte’s AI Nightmare: Top Global Firm Caught Using AI-Fabricated Sources to Support its Policy Recommendations

Close-up of the Deloitte building sign, with greenery in the foreground and a clear sky in the background.

Image: Shutterstock

As we start a new year, 2026, it is a given that artificial intelligence (AI) is going to be the big issue for authors, publishers and the copyright industries generally. The issue of whether it is legal to use copyrighted materials to train AI platforms without the consent of rightsholders will continue to be fought out in courts and legislatures. The use of AI to create content will also continue as an ongoing issue, both the extent to which assistance from AI renders outputs non protectable by copyright and whether content produced using AI is reliable and trustworthy. Deloitte, by revenue the world’s largest consulting firm has just learned that lesson in spades. As well it should.

Would you hire an expensive consulting firm that used AI to supplement its research, didn’t inform you it was doing so, and when it got caught serving up AI-fabricated citations claimed that the false documentation in no way invalidated its policy recommendations and conclusions? I wouldn’t but apparently the Government of Canada has no such qualms, according to a Canadian Press story. We are talking about Deloitte, one of the world’s four largest consulting and accounting firms (the others being PwC, EY and KPMG), a company that charges a premium for its specialized services, and which ought to know better. It’s not as if Deloitte was caught just once with its hand firmly embedded in the AI cookie jar. First it happened in Australia, where the company was forced to reimburse the client, Australia’s Department of Employment and Workplace Relations, for a report that was reportedly “riddled with fake citations, phantom footnotes, and even a made-up quote from a Federal Court judgment.” It also pulled the same stunt in Canada where it produced a report on health care for the Newfoundland and Labrador provincial government. The 500-page report contained at least four citations of research papers that do not exist. These were used, with others, to support recommendations related to recruitment strategies, monetary incentives, virtual care, and the impact of the COVID-19 pandemic on healthcare workers. However, as the Independent, a local digital news outlet reported, Deloitte, having been caught redhanded, stated it is “revising the report to make a small number of citation corrections, which do not impact the report findings.”

What a joke! Of course the fabricated citations impact the report’s findings, as any first year university student knows. Frankly, this is a disgrace and I think Deloitte should be put in the penalty box for six months to a year as punishment, i.e. no government should contract with them until they learn to clean up their act. The irony is that Deloitte advertises itself as providing consulting services to governments to enable them to use AI effectively. Consulting firms are supposed to bring new expertise and perspectives to management problems that governments, in this era of cutbacks, no longer have the resources to solve. They are expensive but provide a quick turnaround for public service managers who don’t have the in-house resources to deal with emerging issues, and who often don’t have the luxury of time to staff up to meet immediate needs. But the dirty little secret is that in many cases the consulting firms apply their cookie-cutter templates to inform their findings whether the template suits or not. They also employ junior staff to do much of the grunt work without, apparently, providing them with adequate supervision or guidance. But even if the labour-intensive task of finding citations to justify the “researched conclusions” of the commissioned report was subcontracted to an AI bot, someone senior at Deloitte signed off on the final product. It didn’t take the client or journalists very long to track down the fabricated citations, so why couldn’t Deloitte have run the same quality check? Because they couldn’t be bothered, I guess.

Despite having been caught, Deloitte may be big enough to shrug this one off, but I sincerely hope they have learned a lesson. Even one false or fabricated citation undermines the credibility of research. As I noted in a recent blog post (Delegating Research to AI is a Risky Proposition: The “Hallucination” Phenomenon-User Beware), “In our rush to embrace AI, many seem to have forgotten the value of human creativity and judgement”. Deloitte has egg on its face, and needs to wear this. A consulting firm is only as good as its reputation, and as far as I am concerned, Deloitte has just put its reputation through the shredder.

This is a cautionary tale, one that has enmeshed not only the world’s largest consulting firm, but various law firms that have been caught citing fabricated precedents. Students would be sanctioned for using AI this way (if they were caught) and academics would suffer major hits to their reputation. Research results and qualifications might be invalidated. If these are the sanctions for misuse of AI, then we should expect no less from entities like Deloitte and its ilk. Let’s hope there are no further AI fabrication horror stories in 2026. (A vain hope, I am sure).

© Hugh Stephens, 2026. All Rights Reserved