actually uses as input: What features or
variables are used in the algorithm? Often those features are weighted: What
are those weights? If training data was
used in some machine-learning process, then you would characterize the
data used for that along all of the potential dimensions described here. Some
software-modeling tools have different
assumptions or limitations: What were
the tools used to do the modeling?
Of course, this all ties back into human involvement, so we want to know
the rationale for weightings and the design process for considering alternative
models or model comparisons. What
are the assumptions (statistical or otherwise) behind the model, and where
did those assumptions arise? And if
some aspect of the model was not exposed in the front end, why was that?
Inferencing. The inferences made
by an algorithm, such as classifications
or predictions, often leave questions
about the accuracy or potential for error. Algorithm creators might consider
benchmarking against standard datasets and with standard measures of accuracy to disclose some key statistics.
What is the margin of error? What is
the accuracy rate, and how many false
positives versus false negatives are
there? What kinds of steps are taken
to remediate known errors? Are errors
a result of human involvement, data
inputs, or the algorithm itself? Classifiers often produce a confidence value,
and this, too, could be disclosed in aggregate to show the average range of
those confidence values as a measure
of uncertainty in the outcomes.
Algorithmic presence. Finally, we
might disclose if and when an algorithm is being employed at all, particularly if personalization is in use, but
also just to be made aware of, for example, whether A/B testing is being used.
Other questions of visibility relate to
surfacing information about which elements of a curated experience have
been filtered away. In the case of Facebook, for example, what are you not
seeing, and, conversely, what are you
posting (for example, in a news feed)
that other people are not seeing.
Technical systems are fluid, so any
attempt at disclosure has to consider
the dynamism of algorithms that may
be continually learning from new
data. The engineering culture must
with five broad categories of informa-
tion that we might consider disclosing:
human involvement, data, the model,
inferencing, and algorithmic presence.
Human involvement. At a high
level, transparency around human involvement might involve explaining
the goal, purpose, and intent of the algorithm, including editorial goals and
the human editorial process or social-context crucible from which the algorithm was cast. Who at your company
has direct control over the algorithm?
Who has oversight and is accountable? Ultimately we want to identify
the authors, or the designers, or the
team that created and are behind this
thing. In any collective action it will
be difficult to disaggregate and assign credit to exactly who did what (or
might be responsible for a particular
error), 21 yet disclosure of specific human involvement would bring about
social influences that both reward individuals’ reputations and reduce the
risk of free riding. Involved individuals might feel a greater sense of public responsibility and pressure if their
names are on the line.
Data. There are many opportunities to be transparent about the data
that drives algorithms in various ways.
One avenue for transparency here is to
communicate the quality of the data,
including its accuracy, completeness,
and uncertainty, as well as its timeliness (since validity may change over
time), representativeness of a sample
for a specific population, and assumptions or other limitations. Other dimensions of data processing can also
be made transparent: how was it defined, collected, transformed, vetted,
and edited (either automatically or by
human hands)? How are various data
labels gathered, and do they reflect a
more objective or subjective process?
Some disclosure could be made about
whether the data was private or public,
and if it incorporated dimensions that
if disclosed would have personal privacy implications. If personalization
is in play, then what types of personal
information are being used and what is
the collected or inferred profile of the
individual driving the personalization?
The model itself, as well as the
modeling process, could also be made
transparent to some extent. Of high
importance is knowing what the model
The main challenge
is to determine
are beneficial but
do not kill usability.