← The AI Daily Brief: Artificial Intelligence News and Analysis
Loading summary
Transcript1 lines
- [00:01]
A
Today on the AI daily brief the case for an AI token tax and maybe the case against it the AI daily brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors, kpmg, Robots and Pencils assembly and zencoder. To get an ad free version of the show, go to patreon.com aidaily brief or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at sponsorsidailybrief AI. At aidailybrief AI you can also find out about everything else going on in the ecosystem. Right now. I am asking for a quick maybe 30 to 45 second set of answers around some ideas for how to make it easier to share AIDB with your teams and get more value out of it that way. And anytime there's something new going on in aidb, you can find it again at aidailybrief AI. Now I am traveling today and so had to prepare this episode in advance. Luckily though, I think this topic was some of the most interesting discourse yesterday, especially after Elizabeth Warren released an op ed in Time magazine about why AI should be taxed. But we are doing a main only type of episode. We should be back with our normal format headlines in domain shortly. Today we're going to talk about the argument for attacks on AI tokens. Now, to be clear, we're also going to talk about the arguments against that, but you better believe that this is a conversation that is just going to increase. Now, one of the things that I feel very strongly is that it is wildly in the interest of the AI industry to not reject out of hand these types of novel policy approaches. If we are indeed entering in such a critically and categorically different period, it follows that policies that have served well enough for many years may simply not make sense in the new context. That does not mean we have to ultimately be in favor of the new policies that get proposed, but I think that the healthiest stance is one of open engagement. Now when it comes to an AI token tax. Specifically, this is a conversation on the rise. It's been around for a while. El Pais, for example, wrote a big piece last November called if AI Replaces Workers, should it also pay Taxes? But it's getting a second wind in a major way right now. Just Yesterday on Wednesday, US Senate candidate from Michigan, Mallory McMorrow released a new comprehensive policy about protecting workers in the age of AI, featuring among other things, a token tax Again, as tempting as it is, especially for the more libertarian minded among you, to reject out of hand any new government policy, I don't think it's particularly hard to tell when someone is coming at the conversation in good faith versus bad faith. And there are a lot of completely reasonable things in McMarrow's policy that I anticipate seeing in other people's platforms and in bills to come. Some of the policies in McMarrow's plan are a little bit more, well, trodden. For example, she's proposing an AI workforce reinvestment fund that requires companies that automate jobs away to contribute pooled resources to a professional apprenticeship program and what she calls a worker centered retraining and upskilling program. However, for our purposes, the more notable one is a token tax, which she calls a modest fee on commercial companies AI usage, ensuring that as AI scales, so do the benefits for working people. Quote, as AI use grows to billions of queries per day, a fraction of a cent charge per token becomes a meaningful, sustainable funding stream for government programs without raising taxes on a single American worker. But it wasn't just a Senate candidate talking about this this week. This is also a growing talking point from people already there in the Senate. In Time magazine on Wednesday, Elizabeth Warren published an op ed called why we need to Tax AI. Again, a lot of it is pretty well trodden territory. Critiques of AI data centers for, quote, jacking up utility bills, concern around an AI financial bubble, but also a growing focus on the implications of AI for what Warren calls our rigged tax code. Warren writes, taxing AI is one way we make sure the winnings from AI benefit all Americans, rather than channeling them only to the wealthy few. If millions of people lose their jobs to AI, we'll need the funds to deliver universal healthcare so those workers are not bankrupted by a visit to the doctor. If AI transforms the future of work, we'll need to invest in free education and apprenticeships and a new jobs guarantee so that all Americans have good paying work. And while workers get back on their feet, we'll need the revenue to bolster unemployment insurance to keep families afloat. The only way we can get there, she writes, is by overhauling our tax code. Now, in the next couple paragraphs, she focuses on tried and true complaints around things like the effective tax rate billionaires pay, basically stuff that has nothing to do with AI itself. However, she writes, rethinking our tax code must also include going to the source. That means taxing AI companies directly, which can start with taxing AI data centers. The majority of AI data centers are controlled or operated by trillion dollar companies. By imposing a reasonable excise tax on the energy used by data centers, families could recoup some of the gains of AI. While America continues to stay competitive in AI race, a well designed tax would focus on the companies that can afford it and scale with AI's impact. The bigger the data center, the more they pay, she continues. We can't be afraid to consider even bigger and bolder proposals to tax AI too, including ideas that sound radical today but may quickly become common sense. If we overhaul our tax code and tax AI, we can use that money to build a country that works for everyone, she concludes. AI was trained on human creativity and intelligence, AI was funded in part by federal investments in scientific research, and AI is powered by data centers that are built on American land and use our shared electric grid. The American people deserve to share in the success of this technology and I'm willing to work with anyone to get it done. Now I will be honest, when it comes to innovation policy, I don't normally find myself particularly aligned with Elizabeth Warren, even if there are plenty of other issues we might agree on. However, I will say in this case, if our options are on the one hand the Bernie Sanders AOC moratorium on data centers, or option B the Elizabeth Warren cut everyone in on the benefits of data centers. I'm certainly more inclined philosophically towards that second position. And recently there have been some surprising voices calling for policies that you might not expect to associate with them. A couple of weeks ago, for example, Mark Cuban tweeted, we should federally tax tokens at the provider level, not a lot less than 50 cents per million tokens. It will accomplish four things at least one it will push the big AI players to optimize tokenization, caching, routing and localization, which will two Reduce energy usage, saving them in energy costs more than what they paid in tax and reducing strain created by the growth in energy consumption which will 3 generate maybe $10 billion a year to start, but over the next 10 years could grow 30x to 100x which will 4 create a source of funding to pay down the federal debt or deploy in response to the things AI brings that we don't expect or don't like. At some point the models will pass it on to customers. Of course that's okay. Customers will have the ability to choose between providers or to do everything using open source models locally. Thoughts and of course he got a lot of thoughts. For some it was the principle of it investor Stephen Sinofsky wrote Imagine a bit tax in 1995. Palmer Luckey disagreed with some of the logic behind Cuban's post, writing, There are already massive economic incentives to optimize, so this is just attacks on American companies that make foreign models and products more attractive, along with creating the infrastructure for government to track all AI usage and punish anyone who doesn't report. Others were a little less generously engaged. Flexport's Ryan Peterson retweeted the post and said claiming a token tax would save AI companies money earns the 2026 dumbest economic thinking Award. Impressive work considering this current year's competition. Another, perhaps unexpected source is DuckDuckGo's Gabriel Weinberg, who doesn't even really dispense with a logical justification for a token tax, but at the end of April basically argued that we should do it and start stashing that money away for a displaced worker fund. He says we should start collecting an AI token tax now and figure out exactly what to do with the funds later, holding them in a true lockbox outside general appropriations. With statutory protection limiting use of funds to supporting displaced workers in the future, we at DuckDuckGo would be willing to support bills to this effect and ultimately pay a token tax presumably presumably collected by the leading AI companies on a usage basis, for example, a 10% surcharge on token charges. The amount would roughly match the 10% employers pay in payroll taxes, which also further reduces the incentive to replace human workers with AI workers. Gabriel is not, however, the only tech leader to suggest such a step. In an interview with Axios last year, Anthropic's Dario Amadei floated the idea of a token tax. Whereas Axios wrote, every time someone used a model and the AI company made money, perhaps 3% of that revenue would go to government and be redistributed in some way. Dario added, obviously that's not in my economic interest, but I think it would be a reasonable solution to the problems. And, axios added, if AI's power races ahead the way he expects, that could raise trillions of dollars. All right, folks, quick pause. Here's the uncomfortable truth. If your enterprise AI strategy is we bought some tools, you don't actually have a strategy. KPMG took the harder route and became their own client Zero. They embedded AI and agents across the enterprise. How work gets done, how teams collaborate, how decisions move not as a tech initiative, but as a total operating model shift. And here's the real unlock that shift raised the ceiling on what people could do. Humans stayed firmly at the center. While AI reduced friction, surfaced insight and accelerated momentum. The outcome was a more capable, more empowered workforce. If you want to understand what that actually looks like in the real world, go to www.kpmg.us AI. That's www.kpmg.usa AI. One thing I keep seeing in enterprise AI companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends, the team's actually shipping. They've picked a lane and they move fast. That's one of the reasons I like today's sponsor, Robots and Pencils. They've gone all in on aws. They're an advanced tier and AWS pattern partner, and they ship production AI coworkers in 45 days. That's led to them doing some of the more interesting work I've seen on AI coworkers. And by that, I'm not talking about chatbots. I'm talking about actual agentic systems that sit inside a business architecture and do real work. That kind of focus matters if you're an enterprise leader trying to get something real into production, or an AWS rep trying to move a customer from interested to deployed. Request an AI briefing at robotsandpencils.com, one conversation with robots and pencils, and you'll know. You know assembly AI for having the most accurate streaming speech to text out there. But they just went a step further and launched a full voice agent API. The idea is simple. One connection, and they handle everything. The listening, the thinking, the speaking. You just stream audio in and get your agent's voice response back. We're talking about things like outbound sales calls that actually qualify leads, customer support that handles complex requests without a script, scheduling, agents that sound like a human assistant. And you can build one in five minutes with one API. And importantly, their streaming model is the best at catching all the stuff that breaks on other voice agents. Things like phone numbers, emails, names and medical terms. And for those of you who are still in experimentation mode, there are no contracts and unlimited concurrency, so you can actually test it out without any friction. Head to AssemblyAI.com brief and try the live voice agent demo right there on the site. No signup needed. So coding agents are basically solved at this point. They're incredible at writing code. But here's the thing nobody talks about. Coding is maybe a quarter of an engineer's actual day. The rest is standups, stakeholder updates, meeting prep, chasing context across six different tools. And it's not just engineers. Sales spends more time assembling proposals than selling finance is manually chasing subscription requests. Marketing finds out what shipped two weeks after it merged, ZenCoder just launched ZenFlow work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools Jira, Gmail, Google Docs linear Calendar notion. It runs goal driven workflows that actually finish your standup brief is written before you sit down. Review cycle coming up, it pulls six months of tickets and writes the prep doc. Now you might be thinking, didn't openclaw try to do this? It did, but it has come with a whole host of security and functional issues which can take a huge amount of time to resolve. Zencoder took a different approach. SOC 2 type 2 certified curated integrations, tighter security perimeter, enterprise grade from day one, model agnostic and works from Slack or Telegram. Try it at ZenFlow free. So what is the argument here from a first principles basis? I think the thrust of Gabriel's argument, which is that we're going to have a bunch of displaced workers that we need to support so we should start collecting it now, is fairly unconvincing, at least on a first principles basis. But there is a certain coherence in the idea that the category shift in who does work from humans to agents creates a shift in the way that taxes work right now that something like a token tax could theoretically solve. Across the OECD, the average tax for a single average worker was 35.1% of labor costs in 2025. Whether you think that's insanely high or too low doesn't really matter. That's the starting point that we're working with now. Suppose, however, that AI agents increasingly perform those same productive tasks, whether it's customer support analysis, medical paperwork, accounting design, or something else. When a human performs the task, the resulting income is taxed through income and payroll taxes. But if an AI agent performs the task, the value might show up as lower costs, higher margins, cheaper services or capital gains, much of which is harder to tax. And even when it is taxed, taxed less labor. By the way, the IMF started arguing about this all the way back in 2024 when they explicitly warned that labor substitution could erode the income tax base if capital income was taxed at less than labor income. So basically, the first principles claim would be something like the tax base should follow the locus of productive capacity. If AI agents become a major class of workers in the economy, some public revenue should be collected from AI work rather than from human work. But why tokens? The reason to tax tokens would Be that in AI systems, tokens are going to be one of the most observable units of AI. Labor providers already meter inference by tokens, meaning that it would be relatively mechanically simple to apply a token tax as a usage based surcharge on top of model inference. Yes, it is an imperfect measure of synthetic labor, but then again, so are hours worked. Tax bases are chosen because they are administratable proxies for something economically important. Theoretically then, you could get some sort of tax neutrality between human and AI laboratory. Imagine option A hiring a person for $100,000, which brings with it payroll taxes, income tax withholding, unemployment insurance, workers comp and compliance costs versus B buying AI agent services for $100,000 worth of inference. Even in this paradigm that we're going into where the AI tokens themselves aren't necessarily much cheaper than humans, but that $100,000 of inference comes with no labor tax equivalent, even if the AI is only slightly better or slightly cheaper. The tax system itself would then push the firm towards automation. Because the human option carries a public finance surcharge, which the A does not. One could argue that that's not free market mediated automation, but instead tax incentivized automation. A token tax, perhaps also paired with a recalibrated or lowered payroll or wage tax, would say in principle, we are not trying to stop automation, we're trying to remove an artificial fiscal preference for replacing taxable humans with untaxed agents. Finally, going a step deeper on the philosophical case, a labor tax implicitly says when a human converts their time or skill or effort into economic output. Society is determined. They are allowed to claim a share of that to fund public goods. That made sense when labor was the main source of production and wages were the main way that ordinary people shared in that growth. But if AI agents perform more of the work, then taxing only human labor ends up taxing the thing that we may want to protect human participation, employment, earned income, bargaining power. A token tax then is a way of saying that the obligation to fund society should be attached to productive capacity, not just human toil. Okay, so that's one version of a first principles argument for this. But as I said, we're not just talking about the case for an AI token tax. We're also talking about the case against it. And there are a lot of cases against it. David Friedman wrote up a bunch of them in a response post to Mark Cuban's proposal. One of the main arguments that he makes is that tokens are a terrible proxy for economic value. One million tokens might be used to generate spam or summarize a novel. They could coordinate a supply chain or create a meme. They could help a student learn calculus, vibe code an app, perform high value legal analysis, or something else entirely. The point being that the economic value per token could vary by orders of magnitude and even more importantly, the purpose of every token consumed is not to produce economic value. Many of the tokens consumed will not be in the service of work. Friedman also points out something that Claude I think interestingly called the tokenizer endogeneity problem. Basically that different providers are going to tokenize the same content differently. Mandarin is going to run two to three times more tokens than English source code is going to be 1 1/2 to 2x times and some low resource languages are going to be 10 to 15x more. A flat per token tax then discriminates between providers and users on a basis completely unrelated to the externality being taxed. Dave writes, this is bad tax design under any framework, but is especially bad when the providers writing the tokenizers are the same parties paying the tax. He also points out that the token tax doesn't have a clean and easy way to deal with the fact that we've seen, as he puts it, a secular 200x annual decline in per token prices that has been the dominant industry trend for two years running. Dave writes, if token prices fall 200x per year and the tax is fixed at 50 cents per million, the tax to price ratio grows 200x per year. A tax that is 5% of a frontier price in year one is 1000% of that same price in year three. Either Congress indexes the tax downward, in which case the revenue collapses or it doesn't, in which case the tax becomes confiscatory at the low end and providers root around it. There is no stable equilibrium where this raises 30 to 100 times more revenue without becoming a different policy entirely. Dave also points out that there's a geographic problem where the tax is levied on providers, but providers can be anywhere and given how much of token consumption is with companies outside the US it means. In effect, he writes, a US per token tax on US Providers is structurally an import substitution subsidy for non US Inference. American enterprise customers serving American end users will increasingly route through foreign domiciled API providers, model hosting platforms, and jurisdiction without the tax or aggregation layers that obscure the underlying provider. Now, interestingly, a lot of the academic work so far on this issue has come to similar conclusions. Brookings sponsored a paper that came out in January of This year called Public Finance in the Age of AI, this tries to deal directly with the possibility of artificial intelligence eventually eroding what they call the two main tax bases that underpin modern tax systems, labor income and human consumption. So this is a group that is acknowledging that in that situation, the burden of taxation will have to shift away from labor. However, what that means in practice is not going to be simple. What they end up doing in that working paper is splitting the AI transition into two stages. Each of those stages have a different set of optimal tax instruments. In fact, they differ dramatically between them. In stage one, which is where labor starts to be displaced but humans are still consuming, they suggest that the right answer isn't a token tax on production, but a consumption tax that captures value where humans actually use services effectively. They're arguing that token taxes in principle are okay, but only at the point of final consumption integrated into VAT and sales tax infrastructure with B2B, use is exempted to avoid some sort of cascading. In stage two, what they call the AGI economy, where AGI is the autonomous producer and consumer, they propose deeper capital taxation on AGI entities. Indeed, one of the big things that they hammer over and over again, both in the working paper itself as well as in their companion blog post, is the importance of distinguishing between intermediate and final use. Effectively, they're arguing that if a token tax applies to business use, research use, manufacturing use, or agentic workflow discovery, it becomes a tax on intermediate production, which distorts productive investment and adoption. Now, this is one of the areas where I find the strongest disagreement with the token tax. In principle, a token tax, as defined, like Mark Cuban does, as just a flat percentage of all the tokens used, creates a significant known ROI bias. Here's what I mean by that. We are already in a period where experimentation with AI use cases is getting costly because there is a greater demand for tokens than the supply of tokens enabled by our compute energy and other infrastructure. Token prices are on the rise. Companies are responding, as you might expect, by placing more severe restrictions on token usage and prioritizing use cases that have known or clear roi. That's going to mean a significant prioritization of what I have frequently called efficiency. AI make customer support cheaper, make analysts produce decks faster, make sales reps write emails faster. It's not that these things aren't valuable, but they're productivity improvements inside an existing model. My very strong contention is that the biggest value from AI is in fact going to be in the new opportunities it unlocks and if we add another layer of disincentive to experimentation, we would be significantly hamstringing the ability for firms in the private market to go out and discover the highest value uses for these tokens. Now, exacerbating that problem is the fact that in no universe would a token tax actually fall equally on everyone. The biggest firms would negotiate discounts, reserve capacity, self host models amortize experimentation over huge revenue bases. Basically, the possibility for this to also entrench incumbents would be just massive. So where all this leads me is that I think that the first principles idea that in a world where the burden of production moves increasingly from human labor to agentic labor might come with some serious changes required to the way we structure the tax base around it. Now, as you've heard me talk about a lot, I also think that the idea of substitution is wildly overblown. So I expect I would disagree on the extent to which the labor tax base gets disrupted, because I think people are still going to have jobs doing all sorts of new things, meaning they'll continue to have wages that can get taxed. But I'm also, generally speaking, sympathetic to the goal of making sure that this transformational technology benefits everyone. I tend to be a big fan of the real politics strategy of buying them off. What do I mean by that? When it comes to data centers and communities, I tend to look for ways that the data centers can be economically valuable in very clear and tangible ways to those communities. Support for jobs and employment is an obvious one, but I think that there are many, many other ways that data center builders and the companies that are going to use them could buy goodwill by subsidizing electricity and other utility costs or creating public infrastructure in some way. Mostly, though, I just think that if the changes are as immense as most of the people listening to this podcast are convinced that they are, we have to be willing to have weird, different, uncomfortable conversations. And so I'm glad we're starting to talk about an AI token tax. I think we can do a lot better, but that's what a good debate is for. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always. And until next time. Peace. Sam.