The bottom of the scale is mostly decorative

You almost never see a 1 because most rating tools are built not to print one. Unscoreable photos get rejected instead of scored, the mapping onto the visible scale treats the bottom third as headroom, and commercial generosity lifts the rest. A rare low number is a design choice, not a compliment to everyone else.

The scale rarely starts where it says it starts

A ten-point scale implies its usable range runs the full length from one to ten, but most tools' actual outputs cluster well above the printed floor - results bunching somewhere between five and eight with the bottom third of the scale essentially unused. That is not the population being uniformly strong. It is the tool's internal mapping treating the low end of the printed scale as a reject range rather than a place ordinary results land, so a submission that would sit near the true bottom of the underlying judgement gets mapped to something closer to a four or a five in the visible number, and the digits one through three sit there mostly as headroom.

Reject-or-score is doing some of the work

Plenty of tools handle a genuinely poor submission - blurry, wrong subject, unusable image - by declining to score it at all rather than by returning a one. A refusal removes the worst cases from the distribution of numbers you ever see, before the scale even gets a chance to use its bottom end. What is left, once the unscoreable submissions are filtered out, is a population that was already somewhat selected, and a selected population compresses toward the top of whatever range remains. Reading a result honestly starts with recognising which parts of it were shaped by filtering rather than by the submission itself, and an unused bottom end is one of the clearest examples of that.

Generosity does the rest

Beyond filtering, most tools are simply built to be pleasant to use, and a tool that regularly hands out ones and twos does not keep users coming back. The same upward pressure shows up in human rating systems: Filippas, Horton and Golden (2019), studying marketplace reputation systems, found ratings "prone to inflation", with raters pressured to leave above-average ratings, which "in turn pushes the average higher". The commercial pressure behind that tendency deserves its own look, since it is the mirror case of the same incentive at the other end of the scale - a tool under pressure to be generous creeps up from below at the same time it hands out tens more freely from above, and the two effects meet in the middle, leaving both ends of the nominal range thin.

What a genuine one would need to mean

For a one to be a real, earned result rather than an artefact of a reject decision, the tool would need a calibration that actually reaches down that far for ordinary submissions - not a filtered population with the worst cases already removed, and not a floor nudged upward to keep the display looking kind. It would need to be treating one as a real position on the scale, occupied by some non-trivial share of whatever gets scored at all, the same way a rubric maximum would need every axis to independently top out before a ten meant what it claims to mean. Almost no tool in this category is built that way. A one you actually see, on a tool that does use its low end, is worth taking more seriously than a one on a tool that never prints anything below a four - the first is the scale doing its job, the second would be almost impossible to produce.

Elsewhere in the category

This is a display-and-calibration question specific to how these tools map an internal judgement onto a printed number - it has nothing to do with the actual model doing the estimating underneath, which is worth understanding as its own layer. It also has no equivalent on the measurement side of the category: a stated length in centimetres does not have a "reject range" the way a rating scale does, it just reports what it reports. A human reviewer sidesteps the whole mechanism differently again - a person responding to a submission is unlikely to hand back the equivalent of a one at all, not because the scale is generous but because a blunt low number is simply not how most people choose to phrase a negative response. On Rate Cock, the visible six-axis breakdown at least lets you see whether a low overall result comes from one axis near the bottom or several sitting evenly in the middle, which is more than a bare total handing out a low number would ever show you.