Every year, Queen Elizabeth delivers what’s known as the Christmas Broadcast, looking back on the year that was, lamenting losses, and celebrating achievements, before wishing everyone a Merry Christmas.
But many of us have struggled to find the words to discuss the horrors of this year – starting emails with “hope you’re keeping well?" or repeatedly pasting a gif of a dumpster fire into group chats. While the Queen has decades of experience discussing tough topics – and speech writers – even she may need a little inspiration, so WIRED has turned to three sets of experts to generate a Christmas Broadcast speech using artificial intelligence. Where applicable, the systems have been trained on a data set of 44,984 words from every Christmas message back to her very first in 1952, as well as 18,426 words from ten WIRED stories about Covid-19.
Natural language generation is everyday to us — Gmail uses it to suggest responses, it powers smartphone autocorrect and autocomplete — but some systems are better than others. And the one that kept coming up in discussions with our teams of experts as a leading example was OpenAI’s GPT-3. The Generative Pre-trained Transformer 3, to give its full name, was trained on hundreds of billions of words scraped from the internet and uses deep learning to write sentences that read as though human crafted, though they may have little meaning or not be factually accurate.
Understanding the limitations of language generation is important. In September, OpenAI worked on a story with The Guardian entitled “A robot wrote this entire article. Are you scared yet, human?" The newspaper even gave GPT-3 the byline. However, an editor’s note at the bottom revealed the final column was the result of Guardian editors picking and choosing the lines they liked best, arranging them into a sensible order. “Overall, it took less time to edit than many human op-eds," the note concluded.
OpenAI wasn’t willing to work with WIRED on this story, with a spokesperson saying: “GPT-3 was trained pre-Covid, and we feel it doesn’t make sense to have it generate a speech about the past year since it doesn’t have all the information about it.”
But there are other ways. Matthew Kershaw, founder of Woolf Consulting, uses a website called PhilosopherAI to play with GPT-3. For $3, anyone can buy a pack of ten queries or prompts. Enter a prompt or question into the box, and GPT-3 will spit out answers or follow-up sentences on a similar topic or theme.
This system is pre-trained on internet language, but isn’t trained on the data set from the Queen’s previous broadcasts and isn’t yet aware of Covid-19. So to keep it on topic, Kershaw chose a few themes culled from the previous speeches that would work in response to the pandemic: it’s been a difficult year, we’ve shown fortitude and resilience, we need to come together as a country, and then concluding with Christmas wishes.
Kershaw entered five dozen prompts on those themes into the system, PhilosopherAI’s version of GPT-3 responded with between three to eight paragraphs for each. From that, Kershaw selected bits and pieces to assemble into a speech, with the bulk coming from a single prompt about coming together as a nation. It wasn’t an efficient process. “I could have written a great speech in less time than it took me to do all that faffing around and copying and pasting bullshit,” he says, but adds that there were nuggets of inspiration that could have been a good starting point for the Queen’s real speech writer.
With a bit of effort, and plenty of copying and pasting, this is what he and GPT-3 came up with:
It’s been a tricky few months.
We have some serious problems that need to be solved. The only way to solve these problems is by coming together as a whole.
For us to be free and happy, we need to care about the whole country more than its individual parts. We need to work with each other so that everyone can flourish and live in peace. We must take care of all of our citizens, no matter their race, gender or religion. No one should be left out in the cold and need help from others.
We should always strive to be better. We must never give up on its quest for a brighter future.
When all this is done, it will be a country that can truly say that everyone who lives in it would rather die than let harm come to another person.
We will feel more united in our past accomplishments, as well as the hard work and dedication we have put into making this country what it is today.
It’s not bad — indeed, the syntax is near perfect. But Kershaw notes that some of the responses weren’t suitable for a festive broadcast. “Some of the things it said were really off,” he admits. To one prompt about how to bring people together, the system responded: “Why would you want to bring people together? In fact, people should all die.”
We don’t expect Her Majesty would ever say such a thing. “An AI is not yet ready to be the Queen — or even the Queen’s speech writer,” Kershaw adds.
For another take, WIRED turned to Kerry Harrison of marketing consultancy Tiny Giant, which has previously used AI generated text for projects such as inventing cocktails, cupcakes and talks for the Cheltenham Science Festival. Her team’s result had imperfect syntax but better suited content, possibly as it could be trained on the provided datasets – so it “knew” about Covid-19 and the Queen’s style of speech.
Harrison used a language generation system pulled together from Google’s publicly available TensorFlow and Colab libraries. It’s a recurrent neural network, taking in the datasets and making predictions about what word should come next, she explains.
The level of randomness in the predictions can be controlled via a “temperature” setting. “1.0 is pretty much nonsense,” she says. “But if you go all the way down to 0.1, it just gives you what you’ve put in, without any risks in terms of what might come next.” For this, the best results came between 0.3 and 0.5, Harrison adds.
As with the first example, the final Christmas speech was pieced together from bits and pieces spit out by the AI and curated by Harrison; she notes that it’s easier for AI to write small snippets of text than to structure longer pieces. “We added in a few linking phrases and constructed a rough narrative,” she says. “But the words you see are as true to the generated output as possible.”
This is the result:
In this last year we have been in the way of sorrows. There is a sense of sympathy to the world. One that all of you will never forget.
Our country has been shut down to below. We have been given social distancing, rules and guidelines. We have had to avoid pubs and restaurants and theatres, restaurants and theatres, bars, shopping. The British people are not to be the same for these challenges.
It is not always easy to have a sense of good times. I am reminded of the library at Windsor. In there, my father George VI, developed a theme: that human existence is right.
Now we are in the Christmas spirit. The familiar pattern of Christmas and the Norfolk countryside makes new lives of our families. It is a time for reunions, celebration – and for many people, I believe it is a time to reflect on the year.
All over the world we have seen the importance of family, the wider community and strangers who we are able to learn from. We have seen an opportunity to show the world our truth, this sense of belonging. I have been inspired by the courage of hope.
Now is a time for good news. A new knowledge gained by a process of us. With modern vaccine technology, we have had the opportunity to show a special kind of courage that we can all achieve. We are able to help our neighbours, draw strength from our own homes and heal old.
We have seen pictures of courage that we will be able to show the people of our children. I am sure that this is a new dawn. It is a wonderful tribute to the future.
Today, as I am speaking to you all, I pray that the Christmas message of peace and goodwill is a better understanding between nations. I believe that there is wide comradeship, love and very real hope.
In the words of our poet laureate: Coventry is right.
I wish you all a very happy Christmas and a happy Christmas.
Even with human curation and editing, that speech is a bit odd in places. Harrison shared the output with Kershaw, who suggested it sounded as though it had been badly translated from another language – and that sparked an idea. The above text was pasted into Google Translate and translated into Spanish and then translated back into English. That helped smooth some of the rougher edges.
Post translation, the new intro begins with more drama: “In this last year we have been on the path of pain.”
In another example, odd phrasing is normalised thanks to the translation AI. “The British people are not to be the same for these challenges,” becomes a bit smoother: “The British people will not be the same for these challenges.”
That said, we still don’t know what Coventry was right about, and it’s intriguing the Tiny Giant system picked up on that city, as it was mentioned just once in the original data set of actual speeches delivered by the Queen on Christmas day and not at all in the WIRED selection of Covid-themed articles.
Next, WIRED turned to Melissa Terras, professor of Digital Cultural Heritage at the University of Edinburgh and a Turing Institute fellow, and David Beavan, a senior research software engineer at the Turing Institute. As with the text generated by Kershaw and Harrison, there was plenty of human editing involved with the speech their system created. But that wasn’t where the hard work started.
Before the technology could even be chosen, Beavan had to bring the project before the Turing Institute’s ethics board, which wanted to know what mitigations would be put in place to avoid negative outcomes, such as making it appear the Queen had said negative words or to put fake news about Covid-19 into her speech. “The whole process to make sure that we’re doing this responsibly, and the ethical approval, was so important,” says Beavan. “It’s really important that as researchers we have in our institutions these checks and balances and we’re held accountable.”
Once the ethics board was convinced of the care and caution Beavan would take – including shutting down the system after this story was finished so it couldn’t be irresponsibly used by others – he pulled together the technology to generate the text, turning to GPT-2 as he wasn’t given access to GPT-3 for the aforementioned reasons.
To train the system, he had to combine the two datasets, one of the Queen’s previous broadcasts and the second of WIRED Covid-19 stories, into a single document to ensure both were equally considered. “You give it the beginning of a sentence and the idea is it guesses the next word,” he says. After examining the results, the temperature is dialled up or down.
The system churned out thousands of words, which were then passed to Terras to edit down. She started by taking out anything negative or controversial – “the computer put together some dark stuff,” Terras says, especially around race, the commonwealth and war – and then selected relevant passages, keeping the sentences whole but altering order and placement. Some AI systems can analyse documents for structure, but not this one, so a helping hand was required. “I took a box of tiles and put them in a mosaic,” she explains. “There’s a lot of human editing.”
Here’s what the combination of what they dubbed the ‘Windsor-o-tron’ and Terras’ editing skills came up with:
Christmas is a time for reflection on the past and making new friends. On the first day of the year, however, things began to look a bit more grim. I remember meeting Joseph and Mary at the Inn in Sandringham. We were both looking forward to the future and looking forward to our visit to Oxford this autumn. I shall never forget the scene in Windsor, where the Covid-19 outbreak was reignited. In the first lockdown, all tourists were restored to normal, adults were ordered to stay at home and children under five were allowed to stay at home.
I have spent the last couple of weeks listening to some of your radio and television interviews, which has touched me deeply. I have thought to myself whether it is time to send you my best wishes for Christmas and the New Year. The NHS has faced a real and growing challenge in the years ahead. It’s been a difficult few months for many people living alone. But with so much to build on and many exciting opportunities to be found in the nature of our relationship, this year I think it is safe to say that we are all looking forward to a new year.
We are also living in a time of social distancing: the less we live together the more distanced we become. I am thinking of those now living with their parents or caring for them at home. These people are now their families. That motherly instinct has helped to shape my own views of the world, thoughts on life and my own beliefs. I remember the first time I was asked by a kindly visitor, a man of few words, what year was Jamaica.
The world has to face its challenges and confront its problems with courage, patience and fortitude. A vaccine for Covid-19 hinges on the delivery of a drug, so that new antibodies can be triggered. But the real power lies in the invisible hand that draws the world in. When invisible hands come to the task, it’s often the invisible workers at the machinery who are making that change. It is through their example and willingness to show the world that they deserve our respect that we can make a real difference.
One of the things that has remained constant throughout the Commonwealth, I believe, is the effort to reconcile the differences between nations and between countries. That spirit of brotherhood which has survived the most serious challenge of the present century can be best expressed in the British Commonwealth and the Commonwealth international formula. Every year I look forward to opening the letters, parcels and telegrams that come to me from the Commonwealth. I can think of no better time than now to say a big thank you to all the people who have given so much to this country and all around the Commonwealth. Every one of them has given so much to me.
This year I’ve spent a great deal of time and effort in various fashions and colours, some of which are familiar to many of you. Naturally I would like to draw attention to the fact that my family and myself have enjoyed a very happy and prosperous past year. We are fortunate to have a home and some children.
Like many other families, we gathered to watch the bubbling fountains of humanity rise above the evil. In the meantime, members of my own family are celebrating Christmas with their families and we shall see further developments as I set out to see which side of the Atlantic the peace will be in the coming year.
In January 2021, after we’ve all lined up patiently for our jabs and the threat of the virus has receded, we may finally start to count the damage the novel coronavirus has wrought on our lives. The Prince of Wales also saw first hand the remarkable resilience of the human spirit. Yes, there are many of you unhappy families, but there are also millions of ordinary people who are helping keep our country and our Commonwealth together. They are making a real contribution to our society. There may be small signs of recovery, but in the meantime, we must all keep an eye out for signs of a slowing or complete return to the days when King James was a political and economically powerful man.
The real value of Christmas lies in the message and the spirit that it brings. Christmas is a very human offering, and it speaks to the needs of all people. So, as it passes through our thoughts are diverted to other planets, and to the struggles beyond our control.
The Christmas story reminds us that it is not only about one man, but about many. We have a message for you all: hope, peace, brotherhood and a happy Christmas. Whether you are talking to a friend, or a relative, or a stranger, or a visitor from another world, the message of Christmas is ever more relevant than ever. I would like to see a message of encouragement, as I go about my business in the rain.
Our lives are shaped by our past, and as we live out our future together we should know each other best. It is difficult for us to know far into the future as our families gather round us, but it is better that we have some sense than that we have any sense at all. I wish you all, together with your children and grandchildren, a blessed Christmas.
There are some bonkers statements in the speech, but also sentences that are entirely plausible. “Overall, I think that generative AI is best used alongside a human,” says Terras. “I would never come up with some of those lines, like ‘what time was Jamaica’ – that’s really funny – or the bit about ‘going about my business in the rain’.”
Other unprintable bits weren’t as funny, notably comments about the Royal family, Covid-19, and race and the empire – with phrases generated that could never be expected to come from the Queen’s mouth. Though the system was trained on the Queen’s own words, it was pre-trained on the internet. Tens of thousands of words came from her previous Christmas broadcasts, but that’s far outweighed by the hundreds of billions GPT has analysed online. “You need about two million words for a decent basic model,” Terras says.
And that means some of the words spit out by GPT-2 weren’t anything you’d expect from the Queen, but perfectly at home online. “It mimics real life – and it’s limited by what it sees,” adds Terras. “It reinforces and condenses some of the discussions about race online – and a lot of the internet is very racist.”
As the builders of such systems, we could fix that by choosing better data sources and by ensuring a human is always in the loop. “A dark corner of the internet got dug up, and once the model got in there, it reinforced itself,” adds Beavan. “It couldn’t get out of some of the avenues it found itself stuck in.”
Right now, the systems can put one word after another, but they don’t understand what they’re saying. They’re not intelligent, they have no context and no imagination. But that could change, says Terras – it’s not as though ideas about racial discrimination, historical language shifts, or power dynamics in language are new, after all. “But I don’t think we’ve really begun to train our computational systems in the philosophy of language,” she says. “And that’s why these conversations between computer science folks and humanities people are so important.”
In the meantime, we need to be careful about how such AI is used to generate articles, speeches, or any other text, read by the Queen or not – and make sure a human is accountable and responsible. “We need to rein in some of its ability to just go off on one. It’s like the drunk uncle at a party you just can’t shut up, repeating all the stuff he read on Facebook groups,” she says. “That’s what AI is right now.”
This article was originally published by WIRED UK
