The key is to have lots of problems nested together, which must be solved in series.
Computers scale a lot better than people, so something which required a human to try to solve a puzzle to get a key, then use that key to decrypt the next puzzle, and so on, probably has better characteristics.
A trusted third party or tamper-resistant hardware is far more practical.
Alternatively, this dongle could contain the necessary private key to decrypt the file instead of the data itself, or another component required to unlock the data a la RSA SecurID.
I would be greatly interested to see relatively secure self-destructing USB sticks.
The first guy was right, it's impossible. Certain very narrow bits of it could be accomplished, but not any real-world goals anyone could have.
The Guardian was operating under a grievous misunderstanding about the nature of the encrypted data, but from my vantage point I don't see that they operated out of intentional malice. If you are distributing data to compliant parties and just want to ensure a tidy cleanup to prevent mishandling or theft, something like this definitely could be useful.
You'd need some sort of physical real-time clock combined with the memory storing the material, which wipes it after a given time. Maybe even a physical medium which degrades over time[3] could work, but that could be foiled by controlling the environmental conditions (inert gas atmosphere to avoid oxidation, cold temps to slow electron migration, etc).
There's a couple of interesting physical-security related links in a comment of mine from the other week: http://news.ycombinator.com/item?id=2932492
My personal approach would be something like providing an incredibly locked-down laptop/netbook (https://grepular.com/Protecting_a_Laptop_from_Simple_and_Sop... would be a good start), but with additional physical security improvements (battery/big caps wired directly to HDD and RAM via a set of tamper switches[1], disabling all IO ports in software and filling them with epoxy / disconnecting internally) You could then wire in an RTC to the same system, as well as perhaps using a GPS receiver to verify the time (Yes, you could jam/spoof GPS signals if you knew to expect them, but that's still raising the bar).
One final approach would be to have some other trusted party/system which remains in your control, and have some challenge/response auth which you can disable/destroy after a fixed time.
To conclude, I can't see any way to build time-limited encryption without some external trusted authority or some trusted physical infrastructure.
[1] Not just physical switches, but as many things as you can come up with: Light sensors, pressure sensors (especially if you can gas-seal the enclosure and keep it at elevated/vacuum pressures), temperature to avoid cooling attacks, resistive/optic-fibre security meshes. Another amusing idea would be to use a GPS receiver to ensure that data can only be viewed from a given physical location[2].
[2] This gets used in _Distress_ by Greg Egan, although I'd thought about it myself long before reading the book.
Edit:
[3] I just remembered about Flexplay (https://secure.wikimedia.org/wikipedia/en/wiki/Flexplay), which was a DVD scheme based on oxidation to time-limit their use as one-shot rentals.