Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
There’s a long history of the US government controlling technology coupled with abuse of that technology, and this pattern repeated itself throughout the 20th century.
This post is a parable. I tell the story of the long history of the government looking to regulate core technology—encryption—as a way of looking at what many are either asking the government to do or more recently what members of Congress are insisting the government must do. Presenting with any technology history, many reflexively assert “this time is different” as they certainly will with respect to AI. The government works in the exact opposite manner. Every situation looks like a past situation, and every action is rooted in precedent. That’s because regulators and the legal system work hand in hand and everything legal is legal via precedent. The question is never first about what we should do now, but what did we do last time and how can we take it further than we got before. That’s because regulation never goes far enough.
AI is different this time because of that very fact. Every past technology used as an example of proof of why we need regulation for technology is talked about as though the present state of regulation has been a long-established precedent. Whether you go through examples from automobiles, airplanes and air travel, drugs and food products, or others the common thread is indeed significant government regulation. But the other thread is just how long those industries were given to mature and define themselves before regulation. Automobiles were not significantly regulated for the first 50 years or so. We had most of the highway system before there was anything more than a minimal test and until “Unsafe at Any Speed” and the aftermath there was little safety testing. Commercial aviation existed for decades before accumulating anything resembling today's comprehensive federal safety regime, an ever-expanding set of regulations and oversight for every aspect of air travel. For example, the original federal transport-pilot license required only 100 hours of solo flight, along with comparatively rudimentary licensing and medical requirements, and you had to bring your own plane. Federal food and drug regulation began early in the 20th century, but for decades it was far narrower than today's regime. safety and efficacy approval did not really approach modern standards until after incidents, many outside the US, in the 1950s and 60s. For example, roughly 3,400 drugs approved between 1938 and 1962 had been approved based on safety without the modern requirement to demonstrate efficacy.
What is happening with AI is a rush to regulate like any industry as ever seen. The technology is immature. The direction of architecture, implementation, and business is hardly established. Even use cases are hotly debated. The rush to regulate is what is unprecedented. It follows the model of the precautionary principle—or essentially pre-regulated to avoid any potential risks, versus defined risks, before there is clarity on what risks are being avoided.
One reason might be because the internet itself was largely unregulated and it could be said we are exiting the experimental phase and now is time for regulation. In practice, the internet demonstrated that most anything that could be regulated was covered by existing laws. Everything from advertising to offering medical advice to selling goods and services are all covered by the same laws and regulations that govern the non-internet. The biggest thing not regulated has been broadly defined as speech. There are many that wish the internet had been defined (in the US) as a common carrier and then the internet businesses it created would have been accountable for what other people said on the platforms. We obviously saw this rise to a significant issue during the pandemic when the administration attempted to thwart what it determined was misinformation, not unlike the kinds of things many presidents have done going back to FDR and radio. It would not be cynical to believe that much of what is happening with AI is simply a continuation of this long effort for the government to find ways to police much more of the public square than they are constitutionally allowed.
AI is closer to a technology mechanism than the internet which was largely seen as a communication tool. In that way it has more similarities to cryptography and encryption which are an underlying mechanism that affords the attribute of secrecy. If there’s one thing that the government has always believed, it is that it has monopoly power on secrecy and secrets. The government historically used the monopoly control of postal mail, phone lines, broadcast spectrum, and microwave transmission/reception including the private companies involved to maintain a monopoly over secret communication. All of this was done to protect national interests. Yet for at least 100 years, there have been countless incidents of the government exceeding legal authority, inventing national security reasons to interfere with private affairs, and in general behaving badly.
The internet has presented a unique challenge for the government to live up to the precedents created by the old systems. The fractures during the pandemic brought that front and center. One lens to use with AI is that it is an attempt to return to past regulatory structures and to frame the 1995-2025 internet as something of an aberration relative to “national security” and private use of computers. The emphasis seen around computer security, potential risk of deaths, China as a global power center, and even bioterrorism, all permit the framing of the AI debate as a national security issue. This is precisely the framing that has been used at every juncture in an attempt to control technology.
It is why I see the risks to preemptive regulation of AI as something more than it might appear. It is why the lessons from cryptography are so relevant. Had the government maintain regulation over cryptography—no matter how ridiculous it seems either technically or Constitutionally—it is reasonable to say we would have halted innovation on the internet. In a relative sense we would have had little by way of the freedom, innovation, and economic growth and opportunity we have on the internet today.
In the legendary book “The Puzzle Palace”—a history of the NSA, National Security Agency—the author details a situation happening in real-time with the government seeking to control modern electronic encryption. The resulting “Data Encryption Standard” (DES) foisted on the private sector was a compromise “championed” by IBM and effectively controlled by the Defense Department. Formally the Depart of Commerce National Bureau of Standards owned the effort and standards process (NBS was the predecessor to NIST) though NSA held a great deal of sway and influence, at the very least. The only thing was that it was almost certainly intentionally weaker than what the NSA was using and weak enough that in a theory outlined by Diffie and Hellman, NSA could build a universal decryption computer for $20M as documented in The Puzzle Palace. In parallel, the State Department, National Science Foundation, and other agencies actively worked to prevent publication of any new research on cryptology and tightly controlled everything from patents to basic mathematics to research funding having to do with the subject.
DES was declared something of a “weapon” by the State Department. Technically the State Department had full authority to ban import/export of any “cryptographic equipment, software, and relevant technical data.” This, at least they believed, provided ample legal room to control the import/export and movement of these DES modules as well as the software and documentation of the product. A DES module was basically an add-in card or stand-alone peripheral for mainframe and minicomputers that handled secret-key encryption, and the private sector standard had a limited key length rendering it breakable by the NSA, should modules fall (illegally) into the hands of the enemy. Along with the hardware, the government agencies extended the reach of import/export control to the knowledge required to make DES. This was controversial and even litigated but the ruling was that the knowledge of the device—and software—was as much a weapon as the device. The final court ruling was technically narrowed to avoid sweeping ordinary scientific knowledge into being banned as the judge was aware of the First Amendment. This is documented in “The Puzzle Palace” which was the first time a broad audience heard of any of this as much of the presented evidence was secret.
A later model of DES was a whole cabinet housing the hardware. Others may have a better recollection, but my recollection is a set of chips on a board that themselves were encased in a cooled translucent housing making photographs impossible. The IBM 4300 where I worked as an intern had one of these units. It was viewed as a scary thing not to be touched.
The knowledge of what the US Government did at this time has been well-established since the 1970s. It goes further back to the pre-War era. The only difference was an alignment of interests of government and researchers, not to mention the complete absence of a private market. The homogeneity of active or tacit approval of government owning a whole technology came from the homogeneity of experience of those involved. Everyone served in the War, had sons who did, or was in some way contributing to it.
Since the 1960s the country has not had this shared experience. This has made “what’s good for the defense of the country, according to the politicians is good for everyone” at best case controversial. It might surprise people to know that the NSA was not formed to intercept all communications everywhere to protect the country and to do so in complete secrecy. The famous executive order creating the NSA within DoD was declassified almost 50 years after its founding. In those 50 years a progression of steps preventing the NSA from intercepting domestic communications came to be.
Except at every juncture loopholes existed so that the NSA could continue to operate inside the US. Some of these were very broad such as any communication between non-US citizens. Others were notoriously broad and later found inappropriate such as any potential criminal activity that might involve other countries or organized crime surveillance. FISA (Foreign Intelligence Surveillance Act) created the first dedicated statutory court-order regime for foreign-intelligence electronic surveillance conducted inside the United States. As we now know that court is unlike any court and by many accounts is far outside what many would call a judicial process.
The energy the government consistently expended to both intercept more and to push the limits of what it knew was permitted legally was, for lack of a better word, relentless. Because authorities and operations of the NSA were extraordinarily secret and the agency maintained an extremely low public profile, there wasn’t even a way to litigate or “sue” as people do today constantly. Even as late as the 1980s when I interviewed for a full-time job in the generally unknown “S Directorate” of the NSA, it was forbidden to acknowledge interviewing or ultimately working there. You identified as a civilian DoD employee.
The key to this evolution is—and really there is no other way to look at this—is that when asked, Americans firmly believe and want the government to protect them from threats foreign and domestic, but are within significant numbers skeptical of overreach, extreme measures, and generally believe that the tradeoffs are not balanced. See for example this 2013 survey after Obama’s administration was caught overreaching surveillance.
The late 1990s brought the internet and some specific technology considerations front and center and for a set of people this has greatly influenced their views of the handling of the role of government as exclusive arbiters, regulators, or managers of technology. Generally, if you did not live through this and it is history, then there is a very good chance you take the most basic parts of the internet for granted, literally HTTPS and end-to-end encryption in messages.
There were many skeptics of DES and the role it was playing, and in the post-Vietnam era where there was a lack of universal consensus on “good” ways to defend the country, a growing number of people saw the needs for citizens and private companies to themselves be secure in their communication…from the government itself in some cases. And many saw the inherent limitations and lack of practicality in private key encryption and aimed for something better. Many involved in DES explained these limitations but NSA working through its “partner” IBM was determined to have a standard that was essentially flawed.
One group of researchers at Stanford hit on a major milestone in cryptography with the invention of practical and computerized public key cryptography. Instead of a secret that practically needed to be shared in person to encrypt data, public key encryption lets each party tell anyone their encryption key while holding their decryption key private and securely. This vastly simplified the process and made it possible for unplanned encrypted communication that was much harder to break.
Many know this by the acronym of the conference that crowds SF every year, RSA. The technology was first discovered as the math by British mathematicians at GCHQ. The intelligence agencies like GCHQ and the NSA had essentially hired tons of math PhDs and prodigies for their work. In 1976, Whitfield Diffie and Martin Hellman openly published the first public description of public-key cryptography, including a practical method for establishing a shared secret over an insecure channel.
It is notable at the time that Diffie also worked with (or perhaps against) the NSA and IBM and was instrumental in attempting to explain the futility—mathematically and fiscally—of securing private key encryption. He noted that as compute power doubled the need for ever longer keys increased. He also took a controversial stance by describing the design as leaving a “trap door” for the government. Hence the short keys :-( Historically, research of the kind by Diffie and Hellman would have been submitted to the Defense Department for pre-clearance authorization. In other words, math was considered a state secret. Unlike many (or most) researchers at the time they were not working under military contract and DoD had little leverage to force them to submit their work to censorship. In fact, the National Science Foundation funded their work and other cryptology work on the west coast.
It is important to note that the NSA (and GCHQ, and KGB) all had monopoly control over advanced and digital cryptography in all forms within their purview. The rise of commercial and private sector computing created a need for commercial cryptography. In theory, DES was a way to meet that demand without relinquishing their monopolies. It was somewhat of a war. The demand was so great that the NSA had little choice but to let the research continue and hope they could control the output.
Perhaps some of this is starting to make sense in the current context of AI?
At MIT during this time the “counterculture” of computing arose. This culture was counter to the past 50+ years of computing essentially focused on computing as a government and military enterprise, especially at MIT. This culture change greatly influenced the rise of personal computing (on the west coast) as well as the free software movement (on the east coast.) Encryption was just part of a large culture shift in computing. This shift was rooted in personal empowerment and distrust of authority in the post-Vietnam era. In my own experience, this aligned 100% with nearly everyone I met as I entered the industry. You can think of these people as a confusing mix of hippie/privacy/libertarian activists. From my perspective this is why today’s AI debates seems to have redrawn the lines of “left” and “right” and brought together people that seemed so apart just 5 years ago.
Given the commercial needs and the incredible utility of this advance you’d think that there would have been cheers all around for this accomplishment. In fact, especially from the government the reaction was anything but that. First there were immediate fears that the NSA would be undermined by enemies of the US using this to communicate and plan against us while doing so with complete impunity from NSA eavesdropping. But as Diffie and others noted, this concern was already the case with DES and simply increasing key length was only a slight delay tactic, even if asymmetrical with commercial interests. But this was just the most obvious first line of “defense” the government used.
Following this, NSA attempted what we’d call “lawfare” today in the hope of using the legal system to shut down the work. In 1977 an NSA employee, officially acting on his own accord and not the NSA, sent a letter to the IEEE warning them that presentations of public key cryptography would potentially result in prosecutions for violations of International Traffic in Arms Regulations (ITAR)—that is an academic paper on an algorithm would be treated a weapon under State Department import/export laws covering weapons (like guns, tanks, and deuterium.) The conference went ahead as planned and as suspected the government probably decided it was a bad look to arrest a bunch of mathematicians at a conference as spies.
NSA then threatened to have their funding pulled and even lobbied NSF to withdraw funding. These were the most prestigious institutions in the world, and it was annoying but hardly difficult for the researchers to find alternate funding. So that didn’t work.
Finally, the fact that Diffie and Hellman had already published their paper the math was already widely known and people not working directly in the area were able to begin to understand and algorithmically create cryptographic solutions. This was historically known as “genie is out of the bottle” era of crypto.
Then in rather short order in 1977 at MIT Ron Rivest, Adi Shamir, and Leonard Adleman invented, proved, and published the first practical algorithm for public key encryption. It had the dual function of protecting communications from eavesdropping while also being able to sign communications preventing modification. Their last names are the RSA we see today in the capability and the company.
What concerned the NSA and broader government most would be the availability and usability of encryption that could be used by criminals or nation-state enemies. Up until this moment, the NSA had a monopoly on encryption and even with the commercially available DES standard they managed to arm-twist IBM into building a standard that the NSA could break at any time. It is also worth mentioning at this time that NSA was routinely involved in domestic surveillance so long as one end of a conversation was a non-US citizen or a US citizen who had recently been overseas, or even if the participants were part of alleged international crimes. The “dragnet” most feared by many but unproven was actually in place as the NSA was able to scoop up vast amounts of communication over common carrier lines and if necessary, even with DES they could decrypt it.
This was about to change in a way no one expected.
In Scientific American, Martin Gardner in his famous column detailed the use of RSA algorithms and encryption for what might soon become very important, electronic mail. Now this was a new kind of idea, not e-mail like we think today. Instead, this was a process where a letter would be “scanned” then transmitted and printed out for delivery by the closest post office. It was going to be the next big thing. It would be important, especially if used by the government, that this would be secret. The race to deliver RSA encryption to consumers was now a thing. Many companies began big plans to deliver this kind of electronic mail.
For the next 8-10 years, RSA worked through commercializing the technology which turned out to be much more difficult than originally envisioned. New microcomputers could not do the necessary math and with IBM dominating mainframes, leapfrogging DES which had been ordained the standard proved difficult. As with a story we hear often, the scientists at RSA brought in a hard-core business leader to be CEO in 1986. Lotus Notes—led by legendary Ray Ozzie—adopted RSA encryption for what at the time was the leading corporate solution for “collaboration” software which included email with their product releasing in 1989. Notes also ran on Windows and Ray was an original Windows Pioneer as one of the technology leaders who bet on Windows early. He later became Chief Software Architect at Microsoft and the person who started Azure and Microsoft’s entry into cloud computing.
Then the internet itself—at the time a small network of academic computers without anything like the world-wide web but rather a network dominated by FTP and some email between university workstations—adopted RSA with a focus on email and three new RFCs 1113-1115, though the patents remained enforced. Microsoft was not even sending/receiving external corporate email on the internet at this time and still using a twice-daily dialup service, precisely because of Gordon Letwin’s (MS-DOS, OS/2) concerns.
RSA, which was patented, offered a RAND license to use its algorithm and protocol. This was decidedly different than the IBM approach of closed-door collaboration with the NSA and Defense Department. This was all done in the open without money or influence. It was an example of the earliest internet efforts in making patented technology broadly available. It is worth noting that at this time the GNU software project was also underway at MIT and projects like GNU C, make, and many utilities were spreading like wildfire across university departments.
In a Kafkaesque moment, the broader US Military command which had historically deferred all cipher work to the NSA (within DoD!) switched to using RSA. The reason was quite simple: they had far too much information to encrypt and managing all those private keys was simply impossible and procuring and managing so much dedicated hardware was too logistically challenging. As we can imagine, the NSA kind of blew a gasket but had little leverage to battle this because they operated entirely in secret. Of course, the NSA focused their objections on the risk that RSA technology would “leak” abroad and be used by our enemies and international crime syndicates. The real concern was that RSA took away the NSA monopoly on being able to both intercept and decode any communication around the world, domestic and foreign.
Many refer to this period starting in 1990 as the “crypto wars” because there were combinations of political battles within the government, technology battles across technical approaches, commercial battles among companies supplying and requiring encryption, and then in 1991 the whole issue erupted into a philosophical battle over the role of privacy and citizen rights versus the government.
This reached a decisive moment at the first RSA Conference held in 1991. The conference was remembered for a panel discussion on the uses of DES and the Digital Signature Standard (DSS) which was a competitor to RSA proposed by the government. The debate centered around the government control of encryption and the obvious limitation of DSS. DSS only solved signing of messages, that is keeping them from being modified in transit, and not the broader protection from being read at all. And just to prove how obvious the NSA was in the process, DSS was also limited in key length, and RSA had no defined limits. The conference panel quickly turned into a broader debate over government control of encryption at all.
This debate is not at all unlike what is going on with open-source AI models today and the attempt by government agencies to develop processes of pre-release vetting of models. These are last-ditch efforts to maintain control over software in a way that maintains a perceived power advantage by the government in the face of alternatives that are already available. It did not make sense for encryption and doesn’t make sense for AI.
If there was a shot heard round the world in the crypto wars, then it would be June 5, 1991, when Phil Zimmermann posted for download PGP or “Pretty Good Privacy” which was a tool to encrypt and sign email messages. It spread quickly on the early FTP and email-based internet via USENET and other BBS services.
I was certainly aware of it and honestly, it seemed kind of fringe to many to be worried about the government snooping on our university email. In reality I was just entering into the world of that weird combination of hippie and libertarian that had already defined the early personal computer era. It is interesting how just a couple of years makes a huge difference in how events are perceived. I see the same thing with AI today where those running the big model companies come from a decidedly millennial optimism about the role of globalization and government that so dominated the political landscape after the fall of the Soviet Union.
PGP was a bit wonky, but it did work. For the next few years, it was integrated into many email clients and transports. It was adopted by IETF and in RFCs. Some sensitive commercial customers were demanding it or compatible approaches. As we were building Microsoft Exchange email for corporations it was clear encryption and signing would use RSA techniques. Microsoft won a bid for email coverage for the entire US military which included requirements for encryption. This also meant all defense contractors and many suppliers of all goods and software to DoD would require matching encryption. Again, the similarities to the influence the DoD has on AI models should be readily apparent. For example, of a model loses its status as a permitted model for DoD then all the DoD supply chain downstream will stop using that model.
There were other encryption skirmishes such as the notorious Clipper chip. There were many instances of international criminals using encrypted email. There were many court cases on these technologies. There was no doubt that the NSA and law enforcement in general felt they had lost. At the same time, many perfectly legitimate end-users and businesses felt much safer from hackers and eavesdropping.
Fairly soon after the release, the US Department of Justice began investigating Zimmermann and his PGP for violations of the US law Arms Export Control Act and the specifics found in ITAR, the International Traffic in Arms Regulations. This was the same rule used to develop and restrict the use of DES. Again, the government has playbooks and precedent, and it used ITAR to restrict distribution and effectively outlaw what amounted to…math.
This investigation went on for years. It was quite incredible. Everyone saw this as a software feature and a public product. But somehow, it was declared to be the equivalent of a Javelin Missile or bioweapon.
In some ways they were messing with the wrong people. Zimmermann and MIT Press published “PGP Source Code and Internals” in 1995 putting the complete source code into an ordinary printed book that could be exported as published material and highlighting the obvious First Amendment problem with treating the same source code as a munition when distributed electronically.
In a form of protest during this time geeks everywhere were sporting RSA encryption code on t-shirts, mugs, and stickers (on paper notebooks.) They were commonly given out by vendors who offered encryption. It was quite a solidarity movement. In many ways the people running open-source models locally on Macs remind me of this energy.
The investigation was ended in 1996. This was just in time for encryption to become a key feature of the new World Wide Web thanks to the needs of private web-based email and internet shopping and the Netscape browser and server.
SSL was designed by Netscape chief scientist, Taher Elgamal, and Kipp Hickman in 1994. It went through a couple of versions before it was robust enough but by 1996 and Netscape 3.0 (and internet Explorer 3.0 as well, but Netscape had 80% share then.) Importantly SSL was also implemented in the Netscape HTTP server, thus offering HTTPS. Microsoft made SSL available roughly at the same time in the same manner with the Microsoft stack, and interoperable. Netscape followed with an open IETF proposal.
From that point forward, there was a march of about two decades before HTTPS became the universal default for browsing. Meaning that your traffic was always encrypted. Many might wonder why it took so long and in practice one of the biggest issues was data center load. We used to have long meetings calculating just how much it would cost Hotmail to run everything as HTTPS. The compute time, the transit of bits in the data center, and more. It was inevitable but it was not without cost. Hardware and optimizations were all part of this effort which was a broad industry-wide initiative. Google as an industry “internet” leader certainly led the way championing HTTPS in major services following the lead of “HTTPS Everywhere” a program started by the Electronic Frontier Foundation in 2010, the leading advocates for internet privacy.
There was also a continuing choice about cryptographic strength. Stronger public-key operations imposed more computational cost, while U.S. export controls forced vendors to ship deliberately weakened cryptography in many international products. This produced different domestic and international versions, additional downloads, negotiated reductions in encryption strength, and interoperability problems. Many vendors during this time, especially those that did a lot of business with the DoD, received subtle “encouragement” to use fewer bits. Often the pitch was “you can save money and handle more data” but really these vendors were identified as throttle points where government actors would like to intercept communications, which might have been legal or not in a market. This type of pressure often led to different bit lengths and approaches around the world. This often manifested itself in software that required extra downloads or global interoperability problems. Office had a significant effort just to maintain the password protection on documents not to mention all the plugins for Outlook/Exchange used around the world.
The encryption store mostly ends here, though we see much of this same tension with end-to-end encryption and messaging products. The desire for government back doors or government preferences in choosing methods to encrypt continues to push on product makers. This happens even though all the code and algorithms are globally understood.
Whether you consider this tension “natural” or “necessary” or “invasive” is all a matter of perspective. For me, it is impossible to look at what is going on with AI today and not think of the journey of encryption. The 50+ year history of the government “reading other people’s mail” and doing so often outside the direct accountability we expect of government. This combined with the pressure and leverage the government maintains over product development even in the face of broadly available source code, software, global alternatives, and broad publishing of algorithms and approaches.
I don’t pretend to have to weigh the costs of decisions required to defend a nation. I do know that the recent sloppiness of products in market have demonstrated that constraining the distribution of the latest technology makes us LESS safe. I hesitate to think of where we would be as a global internet had encryption been constrained or even banned from the earliest days. Even if we had online shopping it would be breachable by pretty much anyone. If our phones were stolen, then any thief would have broad access to our entire lives. I suppose this is just the consumer version of fear mongering over nation-state attacks, but these are everyday occurrences which I would put up against the past decades of less than perfect actions on behalf of States.
Agent shopping represents a fundamental shift in retail
