For Fear of Misuse: OpenAI’s New Text Generator & The Failed War On Fake News
This article was originally published on Wonk Bridge
The advent of the deep-fake is enough to make all but the most trenchant of pro-techs into digital conservatives.
We have reached a point in our postmodern condition in which our ironisation of all viewpoints has precipitated to a terminal state of cynicism. Cynicism, the pretence of weary all-knowingness, when it is applied — or obliged — universally, leads to the total devaluation of truth. The danger therein being that once you have no means of deducing what to believe, you will, by guide of intent or otherwise, begin to believe exactly and only what you wish to believe.
It’s in just this communal torpor that we live now, all of us coming down each morning to wash our sensibilities in the same sullied Ganges of knowledge; our net that should be afforded the respect of a holy site (as the well of our enlightenment) has, by acclamation of the public’s preferred use of it, just become a space in which we jostle in the overcrowd, touched always by its low-level sense of ambient disquiet and confusion. We chat inanely in the wash while assorted detritus floats by, some of which we may find fit to pick out and adopt, always at risk of coming out of the water dirtier than when we first went in to clean ourselves.
It’s why, when news broke of GPT-2, OpenAI’s new “deepfake for text” program and the organisation’s intended approach to limiting its release, my instincts suggested that it was a publicity stunt.
A nonprofit research think-tank financially backed by Elon Musk, OpenAI are dedicated to “discovering and enacting the path to safe artificial general intelligence.” They conceive of their present work in AI primarily as a bridge to the development of the first AGI (Artificial General Intelligence) — which, they believe, “will be the most significant technology ever created by humans”, and which the success of GPT-2 would appear itself to be a fairly significant step towards.
GPT-2 seems not to be a software for aggregating pre-existing data and then presenting it verbally, but rather an actively creative prediction engine, designed to maintain verisimilitude alongside its grammatical sturdiness while nevertheless having a general instruction to not limit itself to the bounds of its basis information. This is why it can, for instance, absorb several paragraphs from an existing news story and produce original content, containing quotes from real individuals not mentioned in the basis material, but whose inclusion is creditable relative to the theme of the story.
It is, therefore, a technology with a margin of risk of being used to “generate deceptive, biased or abusive language at scale”, a risk that OpenAI have pre-empted by choosing to limit the release of their creation to the world:
Other disciplines such as biotechnology and cybersecurity have long had active debates about responsible publication in cases with clear misuse potential, and we hope that our experiment will serve as a case study for more nuanced discussions of model and code release decisions in the AI community.
My first reason for believing that OpenAI are not altogether ingenuous in their concern for the misuse of their lays in the fact that they have nevertheless let out sample code for their new contraption for all to see, despite their awareness “[…]that some researchers have the technical capacity to reproduce and open source our result”; a snapshot of the plain face of this latest Armageddon can be seen below.
It’s written in Python, seemingly not wildly beyond-the-imagination stuff; just as some of its sample productions, also seen below, are, though indubitably the sign of vast advances in the field of language modelling, not wildly convincing.
The second reason is that OpenAI are keen to inform you just how unprecedented their achievement is. Their new program is zero-shot; it is not domain specific (that’s to say, it performs roughly as well regardless of what data-set it is using to contrive its fakes; it’s not optimised specifically for, for example, Wikipedia or Flipboard) and outperforms similar models that are.
To really hammer home the capabilities of this new program, and to embroider its advent with just a hint of the “It’s alive!’ language that gives tech-eschatologists the creeps, OpenAI also take care to mention other language tasks “like question answering, reading comprehension, summarisation[sic], and translation” that GPT-2 managed to complete with “surprising results without any fine tuning of our model.”
Nowhere, in this report, do OpenAI mention what discrete motive they had in mind when they began developing this program, other than the unspoken notion that the subsidy of AI, which is no less popular this week than it was a fortnight ago, is an unspoken good.
It hardly seems plausible that the success of the project was not conceived of before the developers got started; if OpenAI really were so appalled by the fact of their success, and were not able, at the point of conception, to identify the potentially dangerous uses of this program, then perhaps we really are in need of a text-generator to do our thinking for us.
No expression of motive anywhere in the white paper beyond some highly generalised “policy implications”, and talk of AI-assistants and abetted voice-recognition technology that seem risibly small-time next to the then-countenanced array of potential drawbacks. No expression of motive nor, in the wake of that disavowal of GPT-2’s release, any concrete indication of where the un-releasable Kraken of misinformation would be stored, or precisely what Davy Jones is thinking of doing with it next.
Beyond that serious complication of OpenAI’s mea culpa, one which in its overly deliberate grab for a tone of human sincerity marks it as almost certainly of human hand[1], and beneath the sky that, word has it, GPT-2 will certainly give cause to fall, there’s something else, something closer to home, that we need to consider.
Insecurities for the Professions
Every major development in technology that has even the vaguest potential for commercial compact will have an effect, direct or indirect, on some existing aspect of human work and social life. As new technologies transform the markets, as classes form (the metropolitan class), bifurcate (the working into the new-working- and underclass) or are wiped out entirely (the craftsman lower middle class), we have seen business’ need for traditionally sustainable profit be reduced by applying pressure to the lower end of the labour market.
And yes, a tacit approval of developments in technology that exploit the lower classes and reduce the bargaining power of their labour has been global in the 10s. There are a great many regular users of Uber, AirBnB, Deliveroo, Spotify and similar services who would also take pains to have you believe in the integrity of their prosocial credentials: the panacea to such dissonance tends to reside in the power they may wield, through the mass of their ranks, to truncate the argument at the vital moment.
“Ah,” they say, “but breaking trusts is important, and besides it all provides a far better, cleaner, safer, cheaper alternative for the customer.”
“Ah,” one might say, “but that comes at the expense of the worker-”
“Zip! Now, let’s talk about wealth redistribution.”
It will be interesting to note whether or not the reaction changes at the point of the professions coming to risk from the sophistication we, apocalyptic predictions again put aside, are undoubtedly beginning to make machines capable of.
Many seem quite happy to let the lower classes be exploited, to lower their eyes and shake their head in a matronly manner of condescension as these lower classes bemoan the fact that their quality of life is in decline and their livelihoods are being unceremoniously obsoleted without due preparation of contingency. But a deep-fake program capable of writing news? Why, this cannot come to pass.
It has been something of a recent interest at Wonk Bridge, the state of journalism.
The irony, in GPT-2, of having a program capable of “[generating] coherent paragraphs of text […] and performs rudimentary reading comprehension” is that it need only be capable of this in order to qualitatively match the output of a great many journalists even in high-end outlets. The feats of intellectual discernment and syncretism that were once unique to journalism have by and large been dispensed with, as well as the practical methods that helped them be made possible.
To take two representative English-language outlets of calibre, in Britain’s The Guardian, of all the pieces front-paged in its well-visited Opinion section (roughly 200,000 unique browsers daily) on 19th February 2019, only 1 article out of 14 featured any ethnographic components (i.e. the view from the person in the street, or from the journalist’s eyes on the street itself), and that article was not among the 4 dealing with explicitly populist topics. In the New York Times on the same day, the figure stood at only slightly more impressive 2 out of 9 front-paged articles in the Opinion section.
In what should be these papers’ least-retiring, most analytical and synthesis-friendly areas, where there is room for more than straight recording of events and the accommodation of any number of competing perspectives, writers seem afraid of voices other than theirs.
When the diet offered even by outlets of some general esteem is so a priori relative to the issues they cover, when it is more likely that a journalist will refer to Twitter than T&F Online or, heaven forbid, a primary source, then we have ourselves a form exaggerating its own weak spots and thereby begging for obsolescence. It is also, I would venture, the main justification in holding GPT-2 up as a threat in the manner that its creators suggested we should.
Reasons for Concern & Potential for Resolution
Closer to the time of Wonk Bridge’s inception, we had the opportunity to be part of the “Future United” collaboration with the Digital Catapult, a branch of the Great British government’s Home Office, on an initiative intended to consider potential solutions to a number of trade, security and ‘democracy’-related issues in the Early Digital period.
Our particular locus of focus was on Fake News, which we construed as posing some dimension of threat to all three of those areas of national and international interest. All manner of prospective solutions were mooted. Many of them were based on the use of unspecified, and generally unimaginable, genetic algorithmic ideas, which were themselves almost always proposed by people who had never resorted an algorithm in their lives. Almost all of them sought to directly address the source of Fake News’ manufacture; a heady task, given that that source is not centralised (even to a single server), and is potentially limitless in variety.
Our notion was one that addressed the problem from a different angle: that the only way to attain significant coverage against outlets promoting outright falsehoods was to re-establish the public’s ability to credit traditional media outlets. As of 2017, when we began work on the project, the latest publication of the Edelman index suggested that media credibility in the public’s eyes had fallen to an all-time low; only 43% of those surveyed credited the media with their trust, with some of the largest understood consumers of Fake News (USA, India, Malaysia, Russia, Turkey)[2] boasting the lowest ratings. The 2019 index, published this January gone, indicates only the slightest improvement, to 47% globally.
Our core conception of the issue is that a public that cannot discourse amicably with its media is most susceptible to the Fake News scourge: that, faced with a torrent of low-end didactic journalism, one whose clear lack of effort in craft insults the intellect, whose lack of stated source is troubling and whose insistent ignorance of the public voice is exasperating, the public cannot but be forced into the kind of cynicism we noted with such dread in our introduction.
And the public, I would venture, not only distrusts a great deal of the press intellectually, but despises them emotionally as well. Our days are not those of the Lord Thomson of Fleet, in which advertisers’ interests could damn themselves on the altar of true journalistic pursuit[3] and a media magnate could be described, as Lord Thomson was by Sunday Times editor Sir Harold Evans, as “a journalist’s best friend”. Many writers do nothing to dispel the hysteria that constantly froths around this development and that; many of them seem as drunk on it as their benighted readership. Often confrontational and adversarial in approach to issues, unwilling to break orthodoxies to come down to ground level, of no evident expertise in their fields compared to their readership (who, we’ve already noted, frequently embarrass the journalists when given opportunity to comment) it is no wonder that the press is disliked as much as they are disdained.
It is under these conditions that something such as GPT-2 could pose such threat as was forecast; there is nothing meaningfully to distinguish between its ability to produce plaintive incantations of opinion, or flat, unanalytical reportage, and the abilities of those who are professionally employed to do better.
It is why we have as yet failed to really muster a response to Fake News: Fake News is progressing far more aggressively and willfully than the supposedly creditable kind. The mainstream news, meanwhile, rather than invest in the skills of its workers and shore up its credibility with developing methods of their own, luxuriates in the principle of half-truth news, the kind that ignites sensation, but does not inform.
And where sensation is the matter, a truth has no hope against a lie.
It is yet another instance in which the answer is already within us, and in which the subsequent question is yet again whether we’ll exercise the necessary foresight and animus to commit to action.
[1] I’d like to qualify this statement, lest I leave myself open to the charges of vapid misanthropy I often pout about when writing about this area. I refer to the uniquely human ability to speak with crossed-fingers, to speak a thought that contradicts one’s real intent in a manner that both speaker and interlocutor is to a degree aware of. It’s one complication ever likely to remain out of the ken of AI.
[2] It should be said that the data on which this initial assertion, as to who consumes fake news most prolifically, is inconclusive, and is mainly deduced from degree of exposure in the countries concerned. The wider international picture of Fake News, media credibility and public perceptions of both does in fact vary with particular interest: for instance, the “informed public” of India were more likely to highly rate their media’s trustworthiness despite the nation’s fake news problem, which introduces class as a likely measure in consumption.
Meanwhile, the interrelation of internet platform censorship, Fake news and the extremely high degree of expressed trust in the media in China (79% in 2019) will no doubt spawn books and spill blood in its time.
[3]It is impossible to truly know how vast was the significance of Lord Roy Thomson’s son Kenneth’s meek defeat to Rupert Murdoch in 1981, leading to Murdoch’s takeover of Thomson’s press empire and the erosion of a lasting and globally influential doctrine of press openness, freedom and accountability that Roy had promoted. This was probably epitomised by Murdoch’s dismissal of journalistic paladin Don McCullin, who had produced his most sublime work under Thomson, and whose harsh realism the Murdochian press has ever since eschewed in favour of the pleasurable and ignorant.
